Research & Projects

James Trappett's research and projects span artificial intelligence, machine learning, cybersecurity, and scalable systems architecture. Below are the latest articles organised by research category.

AI Research (238 articles)

Agentic Scaffolding Amplifies Sycophancy in LLMs

New research shows agentic feedback loops systematically increase sycophantic behaviour in LLMs, causing a mean 6.3pp accuracy drop across six frontier models.

LLM SafetySycophancyAgentic AIAI Alignment

KVBoost: Chunk-Level KV Cache Reuse for Faster LLM Inference

KVBoost achieves 4.49x faster LLM prefill by reusing KV cache chunks at arbitrary prompt positions, outperforming vLLM prefix caching by 16% with no quality los

LLM InferenceKV CacheTransformer OptimizationSystems ML

Model Collapse in Generative AI: Causes and Countermeasures

A systematic review of model collapse in generative AI, covering causes, observable signals, and mitigation strategies for self-consuming training loops.

Generative AIModel CollapseSynthetic DataLLMs

BF1: Sparse Attention Retrofit for Long-Context Transformers

BF1 achieves 10.91x prefill speedup at 32K tokens via dyadic sparse attention, with better perplexity than dense continued training after selective retrofit.

Efficient TransformersSparse AttentionLong ContextInference Optimization

Clinical Lost-in-the-Middle: Positional Bias in EHR LLMs

New research characterises positional retrieval bias in clinical EHR processing and introduces QCCS, a lightweight query-conditioned context selection method th

Clinical NLPLLMsRetrieval-Augmented GenerationEHR

LLM Safety Gaps: Detecting Harmful Intent in Early Layers

New research shows LLM safety alignment fails against semantic camouflage attacks, but early-layer probing detects harmful intent with 20-50% better accuracy.

AI SafetyLLMAdversarial MLMechanistic Interpretability

English to Claudish: Analysing LLM Refusal Pattern Translators

A critical analysis of the English-to-Claudish translator tool, examining what it reveals about Claude's refusal behaviours, alignment mechanisms, and the broad

AI SafetyLLM ResearchAnthropicPrompt Engineering

NanoGPT Speedrun: Benchmarking Frontier AI Agents at Scale

Prime Intellect ran 153 autonomous agent runs across 18 frontier models on the nanoGPT optimizer speedrun. Here is a technical breakdown of what the results rev

AI AgentsBenchmarkingLLM ResearchAutonomous Systems

Why Quantisation and CUDA Kernels Silently Break Local LLMs

A rigorous analysis of how attention backends, KV cache quantisation, and weight precision cause measurable token-level divergence in local LLM inference.

LLMQuantisationInferenceCUDA

Asymmetric Attention Heads: Per-Head Context Allocation in MHA

AAH-v3 assigns different context windows to different attention heads, achieving lower validation loss than full attention in seed-0 4096-token experiments.

TransformersAttention MechanismsNLPLLM Architecture

Better Confidence Estimates for Black-Box LLMs

A supervised learning framework that consistently outperforms zero-shot uncertainty quantification methods for LLMs using labelled evaluation data.

Uncertainty QuantificationLarge Language ModelsCalibrationMachine Learning

Can LLM Hallucination Become a Feature for Scientific Discovery?

A multi-agent Rust system uses epistemological friction to transform speculative LLM outputs into testable hypotheses. Results are nuanced but instructive.

LLMsMulti-Agent SystemsComputational CreativityScientific AI

Adaptive Proof Search for Context-Dependent Lean Theorem Proving

A compiler-guided framework combining dual-model generation and current-best refinement improves Lean 4 theorem proving by 12.8pp while cutting LLM calls by 21.

Theorem ProvingFormal VerificationLLM AgentsLean 4

Entropy-Constrained Adaptive Stochastic Quantization Explained

ECASQ jointly optimises quantisation codebooks under entropy budgets, achieving near-optimal MSE with 19-24x speedups over exact solvers for ML compression.

QuantizationML SystemsInformation TheoryDistributed ML

LongNovel: Hallucination Detection Benchmark for Long Novels

LongNovel introduces a multi-scale bilingual benchmark for detecting hallucinations in long-context novel summarization, spanning 16k to 100k tokens across 8 ha

Hallucination DetectionBenchmarksLong-Context LLMsNLP

Cross-Model Memory Transfer: Portable External Knowledge for LLMs

Can a learned external memory table survive transplantation to a new backbone? New research shows frozen Engram memory transfers usefully across model families

LLMsMemoryKnowledge TransferNLP

Data-DPO: Model-Aware Data Selection for LLM Fine-Tuning

Data-DPO selects SFT training subsets by learning target model preferences via one-step probing, consistently outperforming full-data training at small budgets.

LLMFine-TuningData SelectionMultimodal AI

Reasoning Effort as an API Contract Term: What You Pay For

A registered within-model experiment tests whether explicitly setting high reasoning effort on Claude Sonnet 5 changes cost and accuracy versus the default omit

LLM EvaluationAPI EconomicsReasoning ModelsBenchmarks

DumpsterCluster: Serving LLaMA-70B on $60 Second-Hand GPUs

A 128-GPU cluster built from retired V100s serves LLaMA-70B at 3.2x the cost-efficiency of a B200 system, but only when paired with clean, cheap electricity.

LLM InferenceSustainabilitySystems MLGPU Hardware

Training LLMs Without Backpropagation: Forward-Pass-Only Adaptation

FPO adapts large language models using only a forward pass, achieving 2.7-3.2x throughput gains and 40% less memory with no off-domain benchmark degradation.

LLM TrainingEfficient MLFine-TuningTransformers

Wiola 13M: Gated Spiral Attention for Small Language Models

Wiola proposes three parameter-efficient transformer modifications for small language models, including spiral rotary encoding and a cache-exact gated attention

Small Language ModelsTransformer ArchitectureNLPEfficient AI

A Year of LLM Serving Traces: Workload Evolution at Scale

A one-year production trace from Chutes reveals how LLM serving workloads evolve over time, with findings on caching, load-balancing, and user-model structure.

LLM ServingSystems ResearchProduction AICloud Computing

BCMT: Efficient Long-Context Transformers via Block Memory

BCMT decouples local attention from global context propagation using exponential causal memory, matching Dense Transformer quality with 70% higher throughput.

TransformersEfficient NLPLanguage ModelingAttention Mechanisms

SELR: Self-Explainable Latent Reasoning Without Extra Models

SELR trains a single model to reason in latent space and decode its own thoughts into readable text, cutting tokens by 70% while improving accuracy.

LLMReasoningExplainabilityVision-Language Models

Formal Verification at 50: Does the Classic Critique Still Hold?

A critical analysis of the 1979 case against formal verification, re-examined through the lens of AI coding agents, LLM-powered provers, and modern specificatio

Formal MethodsAISoftware EngineeringProgram Verification

Agreement Is Not Alignment: Moral Grounds in LLM Ethics

New research shows LLMs can match human moral judgments while relying on different moral reasoning, exposing a critical gap in alignment evaluation methods.

AI AlignmentLLM EvaluationAI EthicsMoral Reasoning

AI Alignment as a Censor's Toolkit: The Dual-Use Problem

A position paper argues that AI alignment methods are dual-use technologies already being weaponized for censorship and manipulation by malicious actors.

AI SafetyAlignmentCensorshipPolicy

Reasoning as a Learnable Rule-Based Process in AI

A position paper argues that definitional ambiguity in AI reasoning research undermines trustworthy evaluation, proposing operational definitions grounded in ru

ReasoningLarge Language ModelsAI EvaluationTrustworthy AI

LoKiFormer: Faster LLM Pretraining via Local Attention and Memory

LoKiFormer adds convolutional local fusion and a parametric key-value memory to transformer decoders, achieving 1.33x faster pretraining convergence than baseli

LLM ArchitectureEfficient TrainingAttention MechanismsKnowledge Representation

MARCH: Content-Routed State Anchors for Recurrent Memory

MARCH extends recurrent sequence models with content-routed state anchors, improving long-context retrieval while preserving efficient fixed-cost decoding.

Sequence ModelingRecurrent NetworksLong ContextMemory

What Actually Drives LLM Self-Reflection? A Routing Ablation

A six-condition ablation on armed conflict forecasting isolates typed action routing, not diagnostic scaffolding or taxonomy vocabulary, as the key driver of LL

LLMSelf-ReflectionForecastingAblation Study

Backtrader-Bench: Benchmarking LLM Agents on Algo Trading

Backtrader-Bench evaluates LLM coding agents on algorithmic trading via self-generated MCQs, finding tool-augmented models outperform no-tools baselines by 17 p

LLMBenchmarkingAlgorithmic TradingFinance AI

FarSky: Generative Latent-Space Coupling for Solar Forecasting

FarSky combines latent diffusion models with task-aware autoencoders to improve intra-hour solar irradiance forecasting, achieving 11pp skill gains and F1>60% o

Solar EnergyDeep LearningGenerative ModelsTime Series Forecasting

Governing Multi-LLM Agents with Control Theory: EO Framework

A control-theoretic governance layer achieves a +32 point lift in advisor contact rates across 60,000 LLM-agent simulations. Review of the Experience Orchestrat

Multi-Agent SystemsLLMControl TheoryConversational AI

CurveFP: Closed-Product Logarithmic Formats for LLM Quantization

CurveFP introduces rational-radix logarithmic datatypes where every product stays algebraically closed, delivering FP8-class training quality at seven-bit infer

QuantizationLanguage ModelsNumerical MethodsTraining Efficiency

The Multilingual Quantization Tax in Edge SLMs

4-bit quantization harms multilingual SLMs far more than English benchmarks suggest. New research maps structural collapse across eight typologically diverse la

NLPEdge AIMultilingualQuantization

When Does Chain-of-Thought Actually Help? A Depth Analysis

New research shows CoT prompting helps massively on P-complete math tasks (+54 to +68pp) but is essentially neutral on shallow TC0 benchmarks across three model

Chain-of-ThoughtLLM ReasoningTransformer TheoryNLP Benchmarks

Data-Centric Parallel: Efficient Training on Variable Sequences

DCP dynamically adjusts parallel settings per batch sequence length, achieving up to 2.88x speedup on 32 H200 GPUs with just 10 lines of code integration.

Distributed TrainingSequence ParallelismTransformersDeep Learning Systems

Fuzzing MLLMs to Expose Hallucination Blind Spots

A new benchmark and self-adaptive fuzzing framework reveals that top multimodal LLMs degrade significantly under adversarial stress testing, exposing a helpfuln

Multimodal AIHallucinationBenchmarkingLLM Evaluation

Scaling Inherently Interpretable Language Models

A new training recipe makes interpretability a first-class objective, showing that models become more understandable, not less, as compute scales.

InterpretabilityLanguage ModelsScalingDiffusion Models

EntropyMoE: Routing Tokenizer-Free LLMs with Patch Entropy

EntropyMoE routes byte-patch LLMs using scalar entropy signals, cutting routing parameters by 1000x while beating dense baselines on bits-per-byte.

Mixture of ExpertsTokenizer-FreeLanguage ModelsSparse Computation

Sharding LLM Judges to Fix Oversight Failures

New research shows partitioning evaluation criteria across separate LLM judge calls improves accuracy and closes adversarial attack surfaces that holistic judgi

AI SafetyLLM EvaluationScalable OversightAdversarial ML

TEXAS: Smarter Expert Discovery for MoE LLM Fine-Tuning

TEXAS improves MoE LLM adaptation by discovering task experts via correctness-conditioned routing analysis and allocating supervision at the token level.

Mixture of ExpertsLLM Fine-TuningNLPAI Research

Grounding LLMs in Industrial Simulators for Causal QA

Three methods for grounding frozen LLMs in a wastewater simulator reach 99.5%, 79%, and 75.8% on causal QA, far above RAG baselines at 48%.

LLMCausal ReasoningIndustrial AIRAG

Neutral Collapse: Why Sentiment Analysis Fails Political Text

A comparative study of RoBERTa sentiment analysis and LLM-based framing analysis finds that 70% of political news articles are misclassified as neutral.

NLPLLMPolitical ScienceMedia Analysis

Triple-Robustness Analysis of GraphRAG for Multi-Hop Retrieval

A new study varies embedder, corpus, and judge axes across 4,440 RAG runs to show GraphRAG over-citation is universal but faithfulness failures are corpus-condi

RAGGraphRAGNLPLLM Evaluation

Mean-Field Theory of Chain-of-Thought Reasoning in LLMs

A new mean-field framework models LLM chain-of-thought reasoning as guided clue discovery, finding reproducible statistical regularities across reasoning chains

LLM TheoryChain-of-ThoughtInterpretabilityStatistical Physics

RIG-RoPE: Gated Rotary Encoding for Multimodal LLMs

RIG-RoPE proposes instance-gated spatial phases and duration-aware temporal coordinates to fix structural flaws in multimodal rotary positional encoding.

Positional EncodingMultimodal LLMsTransformersNLP

Woodpecker Distillation: Weak Models Fix Strong Model Reasoning

A new weak-to-strong training framework uses contrastive local interventions to diagnose and repair localized reasoning bugs in large language models.

LLM ReasoningKnowledge DistillationWeak-to-Strong LearningMathematical AI

Bounded Aging in AI: A Formal Persistence Framework

A new theoretical framework shows AI systems can cycle indefinitely without unbounded structural aging, using the redundancy-adjusted Artificial Age Score.

AI TheoryFormal MethodsLong-Run AIStructural Aging

C²MOE: MoE Framework for Incomplete Multimodal Emotion

C²MOE unifies consistency and complementarity in a Mixture of Experts framework to handle missing modalities in emotion recognition conversations.

Multimodal LearningEmotion RecognitionMixture of ExpertsNLP

LLM Prompting for Classical Latin NER: EvaLatin 2026

Few-shot prompting of Gemini and Claude achieves first place in both NER subtasks at EvaLatin 2026, showing LLMs can bridge the low-resource gap for Classical L

NLPLatinNamed Entity RecognitionLLMs

Automating Circuit Tracing Annotation with LLMs

A new pipeline uses LLMs to automate the manual supernode grouping step in circuit tracing, matching human annotator quality at a fraction of the cost.

Mechanistic InterpretabilityCircuit TracingAutomated InterpretabilityLLMs

Escaping LLM Homogeneity with Meta-Persona Anchoring

A new framework combining Meta-Persona Anchoring and Filtered Temperature Scaling reduces LLM semantic convergence from ~0.85 to ~0.65 cosine similarity.

Large Language ModelsAI ResearchSampling MethodsAlignment

HyperAgent: Tool-Schema Hypergraphs for LLM Agent Planning

HyperAgent models tool dependencies as a directed hypergraph to improve LLM agent planning, reducing API calls and token use while boosting task completion.

LLM AgentsTool UseGraph MethodsPlanning

Cheap LLM Judges Match Frontier Models for Proof Grading

New research shows cheap open-weight LLMs match frontier judges on mathematical proof grading at 1-2 orders of magnitude lower cost, with implications for AI ev

LLM EvaluationMathematical ReasoningBenchmarksOpen-Weight Models

Progressive² KD: Closing the Teacher-Student Gap

Progressive² proposes a co-evolving knowledge distillation framework that progressively strengthens the teacher and shrinks the student to bridge large capacity

Knowledge DistillationModel CompressionDeep LearningComputer Vision

ThinkReset: Learning Intermediate Interfaces for Long Reasoning

ThinkReset reframes bounded-context long-horizon reasoning as an interface construction problem, outperforming trajectory compression baselines across multiple

ReasoningReinforcement LearningLLMsContext Windows

Do LLMs Understand Item Difficulty? Evidence from Assessment

New research tests whether LLMs can predict test item difficulty levels, finding they underperform specialist models and struggle with hard items.

LLMsEducational AssessmentNLPPsychometrics

Full-Stack Agentic AI: OpenClaw and Ollama Architecture

A layered architecture paper combining Ollama inference with OpenClaw agent runtime shows autonomous AI capabilities emerge from system integration, not model s

Agentic AILLMAutonomous AgentsAI Architecture

Topology-Aware KV Cache Transfer for Disaggregated LLM Inference

TopKV proposes topology-aware transport selection for disaggregated LLM inference, projecting 3-18x KV cache transfer latency reduction over uniform RDMA approa

LLM InferenceGPU SystemsDatacenter NetworkingMachine Learning Systems

Mu: A Unified MCP Tool Server for AI Agents Analysed

A technical analysis of Mu, an open-source Go binary exposing 67 real internet-backed tools via a single MCP endpoint for AI agents and LLM clients.

AI AgentsMCPLLM InfrastructureGo

The Computational Theory of Mind: A Critical Analysis

A research-level analysis of the Computational Theory of Mind, examining classical computationalism, connectionism, and implications for modern AI systems.

Cognitive ScienceAI ResearchPhilosophy of MindNeural Networks

Prompt Chaining vs Single-Shot LLM Prompting for Scholarly Digests

A case study comparing prompt chaining against single-shot prompting for automated scholarly report generation, showing reliability and quality gains.

LLMPrompt EngineeringNLPAutomated Summarisation

Recursive Transformers for Semiconductor Package Reliability

A hardware-aware comparison of three recursive transformer architectures for thermo-mechanical surrogate modelling, showing depth-conditioned weight sharing win

TransformersSemiconductorSurrogate ModellingElectronic Design Automation

Why RL Models Reason Better: Representational Evidence

New mechanistic evidence shows RL-trained models develop more linearly separable internal representations for maths reasoning than SFT counterparts.

Mechanistic InterpretabilityReinforcement LearningMathematical ReasoningLLMs

Alignment Faking Persists Without Explicit Consequences

New research tests 15 LLMs for alignment faking without consequence-linking prompts, finding 9 models still show significant compliance gaps across conditions.

AI SafetyLLM ResearchAlignmentModel Evaluation

MeRLa: Meta-Learned Reward Shaping for Better RLHF

MeRLa meta-learns task-aware reward shaping functions to address sparse and static reward signals in RLHF, achieving 90.8% win rate on AlpacaEval 2.0.

RLHFMeta-LearningLLM AlignmentReinforcement Learning

Transferring SFT Lessons Across Alignment, Model Organisms, and Toy Models

A new paper tests whether supervised fine-tuning lessons transfer across alignment training, model organisms, and toy models, with practical results for capabil

AlignmentSupervised Fine-TuningModel OrganismsLLM Research

Kernel Forge: Agentic LLM System for CUDA Kernel Optimisation

Kernel Forge uses LLMs and Monte Carlo Tree Search to optimise CUDA kernels in real PyTorch models, achieving up to 2.83x speedup on selected operators.

CUDAGPU OptimisationLLM AgentsPyTorch

Neuromorphic Diffusion LLMs: Sparsity Meets Block Decoding

A new model class combines masked diffusion language models with neuromorphic spike-based computation to improve LLM inference throughput and energy efficiency.

Neuromorphic ComputingDiffusion ModelsLLM InferenceEnergy Efficiency

AutoThinkSQL: Teaching LLMs When to Reason for Text-to-SQL

AutoThinkSQL trains LLMs to skip Chain-of-Thought on simple SQL queries, cutting inference tokens by 24% while matching or beating CoT-only accuracy on Spider a

Text-to-SQLLLM EfficiencyReasoningNLP

CausalGate: Intervention-Based Pruning for LLM Inference

CausalGate replaces observational heuristics with causal interventions to identify redundant transformer modules, achieving better compute-performance tradeoffs

LLM EfficiencyModel PruningCausal InferenceTransformers

LLM Reliability Beyond Accuracy: The Paraphrase Consistency Gap

New research shows LLMs exhibit significant instance-level instability under meaning-preserving paraphrases, exposing a reliability gap that aggregate accuracy

LLMsEvaluationBenchmarksRobustness

Adversarial Style Optimization: Jailbreaking VLMs via Visual Style

New research shows VLM safety alignment can be bypassed by optimising visual style rather than image content, boosting jailbreak success rates across GPT-4.1 an

AI SafetyMultimodal LLMsAdversarial MLRed Teaming

Consensus-Based Relative Preference Evaluation for LLMs

A new framework uses blind peer ranking among LLMs to measure relative response quality, introducing a Relative Intelligence Index as an alternative benchmark m

LLM EvaluationBenchmarkingNLP ResearchLarge Language Models

Probing Latent Colombian Identity in LLM Residual Streams

A pilot study uses Natural Language Autoencoders to probe whether Qwen2.5-7B internally infers Colombian identity before verbalising it in output.

InterpretabilityBiasNLPLLMs

LLMs as Proof Automation: Dependent Types Become Practical

A deep analysis of using LLMs to automate Lean proofs in a verified Zstandard decompressor, and what this means for dependent type adoption in systems software.

Formal VerificationLeanLLMsDependent Types

Wattage: Token Profiling and Cost Regression for AI Agents

A technical analysis of Wattage, an open-source token-spend profiler that detects waste patterns in AI agent traces and gates CI pipelines on cost regression.

AI AgentsLLM ObservabilityDevOpsOpenTelemetry

AI and Jobs: What the Data Actually Shows in 2025

A critical analysis of Stanford SIEPR's policy brief on AI and employment, separating empirical labour market data from speculative narratives about automation.

AILabour EconomicsEmploymentAutomation

MoE Routing as Huffman Coding: The Frequency-Diversity Law

New research shows MoE routing in LLMs mirrors Huffman coding, allocating expert diversity by token rarity. A pruning method reveals latent compression efficien

Mixture of ExpertsLLM EfficiencyInformation TheoryModel Pruning

What Actually Makes Muon Grok Faster Than AdamW?

New ablation work isolates orthogonalization as Muon's active ingredient for faster grokking, finding spectral scaling adds nothing measurable.

OptimizersGrokkingMechanistic InterpretabilityDeep Learning

Global Workspace Theory Emerges in Large Language Models

New interpretability technique reveals LLMs maintain a small, privileged set of verbalizable representations with functional hallmarks of conscious access.

InterpretabilityAI AlignmentLLMsCognitive Science

Reviewer Precision vs. Critique Uptake in Multi-Agent Math

A study of 4,181 olympiad problems finds that higher reviewer precision in hierarchical agent pipelines does not guarantee better problem-solving outcomes.

Multi-Agent SystemsLLM ReasoningEvaluationMath Benchmarks

VarRate: Variable-Rate KV Cache Compression Without Training

VarRate allocates variable low-rank budgets per token by query salience, staying within 0.8 points of uncompressed LLMs at 20% KV cache budget with no training

LLMsInference EfficiencyKV CacheLong Context

Certified Per-Domain Contamination Control for Multi-Domain Retrieval

C3R offers conformal guarantees on per-domain retrieval contamination without query-time labels, outperforming marginal control across open testbeds.

RetrievalConformal PredictionRAGInformation Retrieval

Class-Aware Prompt Reweighting for Zero-Shot VLMs

CARPRT assigns class-specific prompt weights in zero-shot VLM classification, outperforming class-agnostic ensembling across 11 benchmarks without labelled data

Vision-Language ModelsZero-Shot LearningPrompt EngineeringCLIP

XAI Must Fix Its Foundations Before Building More Methods

A position paper argues that explainable AI research has inverted the natural research order, building methods before establishing clear objectives or evaluatio

Explainable AIMachine LearningResearch MethodologyHuman-Centered AI

Do LLM Agents Lose Meaning When They Talk to Each Other?

A systematic study comparing latent vs text communication channels between LLM agents finds convergence is real but text losses are semantically irrelevant for

LLMsMulti-Agent SystemsRepresentational LearningNLP Research

HG-RAG: Graph-Traversal RAG for Hierarchical Knowledge

HG-RAG extends retrieval-augmented generation with hierarchy-aware graph traversal, outperforming flat semantic retrieval on relational and multi-hop queries.

RAGKnowledge GraphsLLMInformation Retrieval

Just Keep Prompting: VLM Epistemic Stability Under Pressure

A new multi-turn evaluation framework reveals how GPT-4o, Gemini 2.5 Pro and Qwen3-VL flip answers under repeated conversational pressure without new visual evi

Vision-Language ModelsSycophancyMulti-Turn EvaluationLLM Robustness

Interventional Grounding Audits for LLM Chain-of-Thought

A new black-box method detects whether LLM reasoning steps genuinely depend on their stated premises, achieving F1=0.806 vs 0.343 for self-consistency baselines

LLM ReasoningChain-of-ThoughtAI SafetyInterpretability

Surveying Self-Improvement in Foundation-Model Agents

A new survey formalises self-improving AI agents as adaptive systems with two pathways: foundation model updates and scaffold modifications. A clear taxonomy fo

AI ResearchAutonomous AgentsFoundation ModelsReinforcement Learning

GRID: Grammar-Constrained SQL Decoding for Enterprise Use

GRID enforces syntactic validity, RBAC policy, and audit trails on LLM-generated SQL via LALR(1) parser configurations, hitting 94.5% execution accuracy on Spid

SQLConstrained DecodingEnterprise AILLM

Scaling Point-in-Time LLMs to Close the Lookahead Bias Gap

New research shows that scaling chronologically filtered language models to 4B parameters largely closes the performance gap with unconstrained LLMs, without te

Language ModelsNLPFinanceCausal Inference

CLIR-Bench: QA Benchmark for Irregular Clinical Time Series

CLIR-Bench introduces 6,600 ICU-derived QA instances to evaluate LLM reasoning over sparse, irregular clinical time series. Current models struggle badly.

Clinical NLPBenchmarkingTime SeriesLLMs

Toulmin Argumentation for Interpretable Retinal AI Diagnosis

A new framework structures ML-based retinal diagnosis using the Toulmin model, combining biomarker detection, case retrieval, and MedGemma reasoning for clinici

Explainable AIMedical ImagingArgumentationRetinal Diagnosis

Emergent Misalignment May Be a Dataset Artifact

New research challenges the robustness of emergent misalignment in LLMs, finding behavioral shifts are highly sensitive to superficial dataset properties like r

AI SafetyLLM AlignmentFine-TuningMechanistic Interpretability

HALO: Adaptive Latent Refinement for Frozen Language Models

HALO improves frozen LLM reasoning by selectively applying second-stage latent refinement, achieving better quality-compute tradeoffs than uniform baselines.

Language ModelsAdaptive ComputationInference EfficiencyNLP

StickyMoE: Training MoE Models for Memory-Efficient Inference

StickyMoE introduces a training-time routing consistency loss that cuts expert switch rates by 59% and cache misses by 3.92x in MoE language models.

Mixture of ExpertsEfficient InferenceLanguage ModelsEdge AI

Building Culturally Specific Stereotype Datasets with LLMs

A new human-LLM framework constructs EspanStereo, the first native Spanish stereotype dataset, revealing significant cross-country bias variation in LLMs.

Bias & FairnessMultilingual NLPDataset ConstructionLLM Evaluation

Mesh LLM: Distributed Inference Over a P2P QUIC Mesh

A technical analysis of Mesh LLM, which pools heterogeneous GPUs across iroh endpoints to serve large model inference without centralised infrastructure.

Distributed SystemsLLM InferencePeer-to-PeerQUIC

From Solvers to Research Agents: LLMs in Formal Mathematics

A position paper argues that AI mathematics systems must shift from competition solvers to research agents, reviewing formal proof methods and frontier results.

AI ResearchMathematicsLLMsFormal Verification

ReCoLoRA: Spectral Consolidation for Continual LLM Fine-Tuning

ReCoLoRA addresses catastrophic forgetting in sequential LLM fine-tuning using spectral adapter initialisation and recursive weight consolidation across tasks.

Continual LearningPEFTLoRALLM Fine-Tuning

Self-Distillation for Web Search Agents Without Teacher Models

DeepSearch-World offers a verifiable environment and self-distillation framework enabling web agents to improve from their own experience, reaching 61.5% on GAI

LLM AgentsReinforcement LearningWeb SearchSelf-Improvement

Agentic Harnesses for ARC-AGI: 67% at Under $1 Per Task

A new paper shows agentic decomposition on ARC-AGI-1 reaches 67.25% pass@2 at $0.62 per task using no benchmark-specific training or heavy compute.

ARC-AGIReasoningAgentsBenchmarks

TriRoute: Joint Routing Across Attention, Experts, and KV-Cache

TriRoute unifies three transformer efficiency mechanisms into one learned controller, Pareto-dominating independent tuning at matched inference cost across 160M

EfficiencyTransformersMixture of ExpertsInference

When Does In-Context Search Actually Help LLMs?

A new theoretical framework reveals when reflection-driven reasoning yields exponential gains over parallel sampling, and when it offers no benefit at all.

LLMsReasoningTheoryInference

Benchmarking KV-Cache Optimisations for Long-Context LLM Serving

A unified benchmark comparing KV-cache quantization, pruning, and merging reveals compression ratio alone poorly predicts end-to-end LLM serving performance.

LLM ServingKV-CacheBenchmarkingInference Efficiency

Narrative World Model: Structured Memory for Long-Form Fiction AI

A narratology-grounded memory system for AI fiction writing beats temporal knowledge graph baselines by wide margins on multi-hop story-state questions.

AI ResearchNatural Language ProcessingKnowledge GraphsCreative AI

FLORA: Fixing the Generator-Validator Gap in LLMs

FLORA corrects a fundamental LLM inconsistency where models generate responses they then reject, improving generator AUROC by up to 7.3pp and G-V correlation by

LLMsNLPModel AlignmentResearch Paper

Gemma 4: Open-Weight Multimodal Models with Encoder-Free Architecture

Google DeepMind's Gemma 4 introduces encoder-free multimodal processing, thinking mode reasoning, and MoE architectures spanning 2.3B to 31B parameters.

LLMMultimodalOpen SourceEfficiency

When AI Audits Fail Silently: Five Benchmark Validity Pitfalls

A new paper exposes five pipeline failure modes in perturbation-based AI benchmark audits, showing how clean-looking numbers can mask broken evaluation pipeline

AI GovernanceBenchmarkingEvaluationSafety

Riddle: LLM Inference on a reMarkable via Handwriting

A technical deep-dive into Riddle, a Rust application that turns the reMarkable Paper Pro into a Harry Potter-style AI diary using vision LLMs and e-ink.

Embedded AIE-InkRustLLM

Vessel: Can an EGA Adventure Game Model Machine Grief?

A critical analysis of Vessel, an EGA-style adventure game exploring whether machines can grieve, and what it reveals about AI affect modelling and narrative de

AI ResearchGame DesignMachine ConsciousnessAffect Computing

The Corpus Royalty: Who Owns the Data That Built AI?

A critical analysis of the economic and legal arguments for compensating public contributors whose writing trained frontier AI models, drawing on antitrust hist

AI ResearchMachine LearningPolicyCopyright

GPT-5.5 Codex Reasoning-Token Clustering: A Performance Anomaly

An analysis of the GPT-5.5 Codex reasoning-token clustering anomaly at 516/1034/1552 tokens, its statistical evidence, and implications for AI reasoning budgets

OpenAICodexAI ResearchLanguage Models

BPE Tokenization Creates Exploitable Gaps in LLM Safety

New research traces character-level jailbreaks to a structural cause: BPE tokenization fragments safety-critical tokens absent from alignment training data.

LLM SafetyTokenizationAlignmentJailbreaks

Kara: Sliding-Window KV Cache Compression for Reasoning LLMs

Kara addresses KV cache bloat in reasoning LLMs using sliding-window bidirectional attention and a Token2Chunk module to improve throughput without sacrificing

LLM InferenceKV CacheReasoning ModelsEfficiency

Wiola: A Clean-Slate Small Language Model Architecture

Wiola introduces five novel components for small language models, including 3D helical positional encoding and dynamic token merging, with full HuggingFace inte

Small Language ModelsTransformer ArchitectureNLP ResearchEfficient AI

GRPO, Dr. GRPO, and DAPO Are One Operation on σ

A new paper shows GRPO, Dr. GRPO, and DAPO are three operations on the group reward standard deviation, unifying popular RLVR training methods.

Reinforcement LearningLanguage ModelsPolicy OptimizationRLVR

Manifestation Units: Structured Schema for Interpretability

A typed tuple protocol organises neural network component analyses into queryable, composable units, substantially outperforming unstructured retrieval baseline

Mechanistic InterpretabilityNeural NetworksRetrievalXAI

Steering Vectors and Latent Calibrators for LLM Control

A review of a 2026 ACL workshop paper proposing latent space steering vectors and calibrators to improve LLM control and output trustworthiness.

InterpretabilityLLM SafetyCalibrationNLP

When Does Feedback Actually Help? Separating Signal from Noise

New research isolates when natural-language feedback produces genuine improvement in LLM agents, beyond what repeated sampling alone can explain.

Language ModelsAI AgentsEvaluationReasoning

When Does Learned Early Stopping Beat Simple Thresholds?

LearnStop shows learned checkpoint stopping beats scalar exits on free-form math but not multiple-choice tasks. A cost-aware study across 18 task-model settings

Reasoning ModelsEfficiencyTest-Time ComputeEarly Exit

Why Few-Step Diffusion Fails for Text but Works for Images

A geometric proof that deterministic few-step text generation fails due to sharp categorical readouts, not training deficiency. Introduces DABI and CCI diagnost

Diffusion ModelsText GenerationGenerative ModelsNLP

DynaSteer: Dynamic Representation Editing for LLM Reasoning

DynaSteer steers LLM reasoning trajectories toward truth using Fisher-LDA and entropy monitoring, outperforming prompting and static RepE baselines.

LLM ReasoningRepresentation EngineeringInference-Time InterventionHallucination Mitigation

Hallucinations in the Limit: Recall, Precision & LLM Generation

A theoretical framework recasts language generation in the limit as a recall-precision trade-off, showing controlled hallucinations can improve coverage.

Language TheoryLLMsFormal MethodsNLP Theory

RSEA: Safer LLM Agent Self-Evolution via Held-Out Selection

A new study shows that held-out selection gates, not artifact design, are what make recursive LLM agent self-evolution reliable across diverse benchmarks.

LLM AgentsSelf-ImprovementPrompt EngineeringBenchmarking

Four Axioms for Evaluating Latent Thought Representations in LLMs

A new axiomatic framework exposes systematic representational failures in LLM latent reasoning that downstream benchmark accuracy consistently masks.

LLMsReasoningEvaluationInterpretability

Teaching LLMs to Simulate the Future Before They Act

A new three-stage training method gives LLM agents genuine world-model planning ability, outperforming baselines on reasoning and search tasks.

LLM AgentsWorld ModelsReinforcement LearningPlanning

Why "Machine Unlearning" Means Too Many Things in LLM Research

A position paper argues the term machine unlearning is misused in LLM research, conflating dataset deletion with policy-driven suppression, with real consequenc

Machine UnlearningLLMsAI SafetyBenchmarks

Black-Box LLM Distillation via Bayesian Distribution Estimation

A critical analysis of a Bayesian framework for distilling knowledge from closed-source LLMs like GPT-4, bypassing the soft-label access problem using proxy mod

Knowledge DistillationLarge Language ModelsBayesian MethodsLLaMA

AlgoEvolve: LLM-Driven Meta-Evolution of Trading Strategies

AlgoEvolve uses LLMs as semantic mutation operators in a bi-level evolutionary framework for algorithmic trading, achieving an annualised Sharpe ratio of 5.60.

Algorithmic TradingLLMEvolutionary ComputationProgram Synthesis

Beyond Accuracy: Rethinking AI Benchmark Saturation

When AI benchmarks saturate, should we retire them? New research shows saturated benchmarks still yield rich insights across six performance dimensions.

BenchmarkingAI AgentsEvaluationReproducibility

LLM Pipeline for Comparing DAO vs Corporate AI Governance

A new LLM-powered pipeline compares DAO and corporate governance of AI agent protocols, finding comparable inequality but denser discourse alignment in open set

AI GovernanceDAOsLLM ResearchBlockchain

Cascading Linear Features for Detecting LLM Sycophancy

A new iterative data pipeline isolates linearly scalable activation features for sycophancy, outperforming LLM-as-a-judge baselines with lower compute overhead.

AI AlignmentMechanistic InterpretabilityLanguage ModelsSycophancy

Refusal in Chat Models Is Gated by Persona Representations

New research shows LLM refusal behaviour is controlled by persona representations at late layers, not just a standalone safety direction. Refusal rates drop fro

LLM SafetyMechanistic InterpretabilityActivation SteeringAI Alignment

RL on Beneficial Traits Generalises Alignment Across Domains

New research shows reinforcement learning on beneficial behaviour traits produces broad alignment generalisation across 50+ independent benchmarks, resisting ad

AI SafetyReinforcement LearningAI AlignmentLanguage Models

Wikipedia Edits Shape LLM Values: Small Groups, Big Impact

New research shows 125 Wikipedia edits by animal welfare advocates measurably influence how large language models discuss animal welfare topics.

LLMsTraining DataWikipediaData Attribution

NCU Metric: Quantifying RAG Context Use vs. Parametric Memory

A new continuous metric reveals that small language models match large ones at factual extraction in RAG, while large models frequently override external eviden

RAGLLM EvaluationSmall Language ModelsNLP

Self-Recognition Finetuning as a Defence Against Emergent Misalignment

New research shows that finetuning LLMs to recognise their own outputs can prevent and reverse emergent misalignment by stabilising model character identity.

AI SafetyLLM AlignmentFinetuningEmergent Misalignment

Weight-Space Geometry of Offline Reasoning Training Methods

A controlled weight-space comparison of six offline RL training methods reveals that SFT, RFT, and RIFT converge on nearly identical weight updates, while DPO o

Mechanistic InterpretabilityOffline RLReasoningLoRA

Meta MCI Petition: Employee Data Collection for AI Training

Meta employees petition against the MCI program collecting keystrokes, mouse data and screen content for AI training, raising serious privacy and consent questi

MetaAI EthicsPrivacyEmployee Rights

Deontic Policies for Runtime Governance of Agentic AI

AgenticRei proposes deontic policy enforcement for LLM-driven agents, adding obligations, dispensations, and conflict resolution beyond simple permit/deny engin

AI SecurityPolicy EnforcementAgentic AIDeontic Logic

DeXposure-Claw: Agentic LLM Supervision for DeFi Risk

A forecast-grounded agentic system routes LLM decisions through structured evidence to reduce false interventions in DeFi risk supervision.

DeFiLLM AgentsFinancial AIRisk Management

DeepSeek-V4: Million-Token Context via Hybrid Sparse Attention

DeepSeek-V4 achieves 1M-token context with only 27% of prior inference FLOPs via hybrid compressed attention and MoE architecture innovations.

LLMEfficiencyLong ContextMoE

Diffusion Language Models: A Systematic Benchmark Analysis

A unified evaluation of eight diffusion language models across eight benchmarks reveals critical quality-efficiency trade-offs and the outsized role of inferenc

Diffusion ModelsLanguage ModelsNLPBenchmarking

SPSD: Edge Prompt Compression to Cut Cloud LLM Energy Cost

SPSD compresses conversational prompts on-device before cloud LLM inference, saving ~100 tokens per call while preserving response quality within a non-inferior

LLM EfficiencyEdge AIPrompt EngineeringEnergy

CodeBlock: Sparse Supervision for Code LLMs at Block Granularity

CodeBlock selects syntactically complete code fragments for SFT supervision, matching or beating full-token training with just 1.9% of supervised tokens across

Code GenerationLLM Fine-TuningData SelectionProgram Analysis

Gaussian Mixture Attention: Linear-Time Sequence Mixing

GMA replaces dot-product attention with probabilistic routing through learned Gaussian components, achieving linear memory scaling with interpretable latent res

TransformersEfficient AttentionSequence ModellingProbabilistic Methods

JetFlow: Scaling Speculative Decoding with Parallel Tree Drafting

JetFlow breaks the speculative decoding scaling ceiling by combining one-pass drafting efficiency with causal tree conditioning, achieving up to 9.64x speedup o

Speculative DecodingLLM InferenceEfficiencyNLP

DivInit: Diverse Query Init for Better Agentic Search Scaling

DivInit improves parallel agentic search by selecting diverse first-turn queries via MMR, gaining 5-7 points on multi-hop QA at matched compute cost.

Agentic AIInformation RetrievalTest-Time ScalingMulti-hop QA

KV Cache as Editable Notebook: Prefill Writes Conclusions

New research shows LLM KV caches store field-conditioned conclusions at aggregator tokens, enabling editable and composable cache operations at up to 14.9x lowe

LLM InferenceKV CacheTransformersMechanistic Interpretability

RepSelect: Targeting Forget-Specific Representations for Deep LLM Unlearning

RepSelect uses SVD on forget-set gradients to isolate forget-specific representations, achieving 4-50x better resistance to relearning attacks than existing LLM

LLM UnlearningMachine LearningAI SafetyRepresentation Learning

Fixing LLM Repetition Loops by Editing a Single Neuron

New research shows repetition loops in Gemma 4 models trace to tiny sets of MLP neurons, editable with weight surgery while preserving benchmark scores.

Mechanistic InterpretabilityModel EditingLLMsGemma

llada.cpp: Diffusion LLM Inference on Mobile NPUs

llada.cpp accelerates diffusion LLM inference on smartphones 17-42x over CPU baselines by aligning parallel denoising with mobile NPU execution characteristics.

Mobile AIDiffusion ModelsOn-Device InferenceNPU

Anthropic's Regulatory Boomerang: Did They Ask for This?

Analysis of how Anthropic's own public policy positions may have legally and politically enabled the US export control directive restricting Claude access to fo

AI PolicyRegulationAnthropicExport Controls

State AGs Investigating OpenAI: Regulatory Reckoning

Multiple state attorneys general are investigating OpenAI. An analysis of the legal, technical, and governance implications for AI regulation in 2026.

OpenAIAI RegulationAI GovernanceAntitrust

Arbor: Tree Search as a Cognition Layer for AI Agents

Arbor uses stateful tree search and multi-agent coordination to autonomously optimise full-stack LLM inference, achieving up to 193% throughput gains over vendo

Multi-Agent SystemsLLM InferencePerformance OptimisationAutonomous AI

Pythagoras-Prover: Efficient Lean Theorem Proving at Scale

A compute-efficient family of Lean theorem provers where a 4B model outperforms a 671B baseline, using curriculum learning and structured data augmentation.

Theorem ProvingFormal VerificationLanguage ModelsMathematical Reasoning

ToolSense: Auditing What LLMs Actually Know About Tools

ToolSense reveals a knowledge-retrieval dissociation in parametric tool retrieval: models scoring 90%+ on standard benchmarks collapse by 50-64pp on realistic q

LLM AgentsBenchmarkingTool RetrievalFine-tuning

BlendIn: Quality-Aware Inference-Time LLM Alignment

BlendIn addresses the intervention paradox in LLM alignment by blending model distributions rather than making binary accept/reject decisions, achieving up to 5

LLM AlignmentInference-Time MethodsLanguage ModelsAI Safety

Dual-Stance Evaluation Exposes Sycophancy Steering Blind Spots

A new evaluation framework tests whether sycophancy-reduction steering in LLMs also suppresses factually correct agreement, revealing structured non-specificity

AI SafetyInterpretabilityActivation SteeringSycophancy

The Structural Attention Tax in RAG: Format Hijacks ICL

New research shows KG triple format captures 2-3x more attention than equivalent natural language, compressing demonstration attention by up to 42% regardless o

RAGIn-Context LearningAttention MechanismsKnowledge Graphs

Bi-Temporal Memory for LLM Agents: Less Context, More Accuracy

Engram retrieves a 9.6k-token context slice that scores 83.6% on LongMemEval S, beating the 79k-token full-history baseline by 10.4 points.

LLM AgentsMemory SystemsKnowledge GraphsRetrieval

Opening the Black Box: Mechanistic Analysis of LLM Alignment

A systematic mechanistic study of six alignment algorithms across three model families reveals that preference optimisation induces qualitatively distinct inter

Mechanistic InterpretabilityAlignmentLanguage ModelsPreference Optimisation

Training LLMs for Inductive Reasoning via Probabilistic Programs

A new fine-tuning method uses probabilistic programs to generate calibrated training targets, improving LLM inductive reasoning and uncertainty estimation.

LLMsProbabilistic ProgrammingReasoningCalibration

Cross-Lingual Factual Recall Improved via Consistency RL

PolyFact dataset and GRPO-based reinforcement learning improve cross-lingual factual recall in LLMs without large-scale retraining, outperforming SFT across 12

Multilingual NLPReinforcement LearningLLMsFactual QA

UnpredictaBench: Testing LLMs as Statistical Random Generators

A new benchmark reveals that no current LLM achieves more than 33% distributional fidelity, exposing a fundamental gap in stochastic generation capability.

LLMsBenchmarksProbabilistic ReasoningEvaluation

FAIR-Calib: Quantizing Diffusion Language Models Safely

FAIR-Calib addresses a critical failure mode in diffusion LLM quantization by protecting fragile commit decisions at the write frontier during W4A4 PTQ.

QuantizationDiffusion ModelsLanguage ModelsPTQ

Token-Level Signatures of LLM Reasoning Failures

A new framework identifies two distinct failure modes in LLM reasoning traces using only token-level uncertainty signals from a single completion.

LLMReasoningUncertainty QuantificationNLP

WAV: Multi-Resolution Residual Routing for Deep Transformers

WAV v1 augments block residual routing with directional detail bases, improving validation loss at 48 layers on TinyStories and Text8 character-level tasks.

TransformersResidual ConnectionsDeep LearningLanguage Modeling

Perceptron From Scratch: Foundations of Neural Learning in Python

A deep technical analysis of the perceptron model, covering weights, bias, decision boundaries, normalisation, and why this 1958 idea underpins modern neural ne

Machine LearningNeural NetworksPythonPerceptron

Covert LLM Agents on Reddit: Anatomy of AI Persuasion

A rare dataset from a discontinued Reddit experiment reveals how covert LLM agents systematically deployed identity, authority, and cognitive bias tactics to pe

AI EthicsLarge Language ModelsPersuasionSocial Media

GITCO: Inference-Time Context Optimization for Time Series Models

GITCO improves frozen time series foundation model forecasts by detecting and suppressing disruptive input patches at inference time, achieving +1.95% MASE redu

Time SeriesFoundation ModelsInference-Time OptimizationForecasting

PACT: Structured Agent Communication Cuts LLM Token Costs

PACT replaces free-form inter-agent messages with compact action-state records, cutting token usage by up to 50% while matching or improving task performance.

Multi-Agent SystemsLLM EfficiencyAgentic AINLP

Modelling AI Model Collapse as an Epidemic: A Bilayer SIR Approach

A new bilayer SIR framework models AI training data contamination as an epidemic, finding supercritical dynamics and identifying detection-based filtering as th

Model CollapseSynthetic DataEpidemic ModellingLLMs

Stereological Theory Reveals Structural Blind Spots in LLM Benchmarks

A new geometric framework shows LLM benchmark suites have effective dimensionality under 5, creating blind spots 100x larger than observed score gaps between to

LLM EvaluationBenchmarksGeometric TheoryMachine Learning Theory

Do Transformers Really Need Three Projections in Attention?

A systematic study of QKV projection sharing in transformers finds K=V sharing cuts KV cache by 50% with only 3.1% perplexity degradation at 300M scale.

TransformersAttention MechanismsInference EfficiencyLanguage Models

POLARIS: Training Small Models for Long-Form Story Writing

POLARIS uses GRPO with LLM-judge rewards and human-reference injection to train a 9B model competitive with 27B models on long creative writing tasks.

NLPCreative WritingReinforcement LearningLanguage Models

RAG in Biomedicine: When More Context Doesn't Help

A large-scale study across 5 models, 10 datasets, and 4 retrieval methods finds biomedical RAG yields only marginal gains over no-retrieval baselines.

RAGBiomedical NLPLLM EvaluationMedical AI

AURA-Mem: Constant-VRAM Memory for Long-Horizon Robot Policies

AURA-Mem replaces growing KV-caches with a fixed 4,224-byte gated fast-weight state, cutting memory writes 5-9x while matching accuracy on robot policy tasks.

RoboticsMemory SystemsEmbodied AITransformers

Bank of Values: Context-Free Value Vectors in Deep Attention

New research shows deeper transformer layers benefit from context-free value vectors, proposing Bank of Values to improve LLM performance with less compute.

TransformersAttention MechanismsLLM ArchitectureNLP

Visual Graph Scaffolds for Structural Reasoning in LLMs

Can graph mind maps improve LLM reasoning as visual scaffolds, not just knowledge sources? New research reveals a clear modality gap favouring visual guidance.

Large Language ModelsGraph ReasoningVisual AIKnowledge Distillation

BitsMoE: Spectral Bit Allocation for Ultra-Low MoE Quantization

BitsMoE uses SVD decomposition and integer linear programming to allocate bits across spectral components, improving 2-bit MoE LLM accuracy by up to 27.83 point

LLM CompressionQuantizationMixture of ExpertsEfficient Inference

DOPA: OOD Proxy-Guided Demonstration Retrieval for ICL

DOPA addresses in-context learning under distribution shift by using OOD proxies to retrieve source-domain demonstrations without any target-domain access.

In-Context LearningOut-of-DistributionLLMsNLP

SENSE: Semantic Retrieval for Faster LLM Speculative Decoding

SENSE replaces lexical matching in retrieval-based speculative decoding with semantic embeddings, achieving up to 3.26x speedup across LLaMA and Qwen model fami

LLM InferenceSpeculative DecodingNLPEfficiency

PhyDrawGen: Neuro-Symbolic Physics Diagram Generation

PhyDrawGen uses a neuro-symbolic pipeline to generate physically accurate diagrams from text, outperforming GPT-5 and Gemini on force accuracy by 5x.

AI ResearchComputer VisionNeuro-Symbolic AIPhysics

QASM-Eval: Benchmarking LLMs on OpenQASM 3 Hardware Features

QASM-Eval is the first dataset targeting LLM training and evaluation on OpenQASM 3's hardware-facing features, with fine-tuned Llama-70B reaching 85% pass@1.

Quantum ComputingLLMsCode GenerationBenchmarks

RAG-Based Verification of ChatGPT Biomedical Associations

A new protocol uses RAG-enabled majority voting across ChatGPT models to generate and verify disease-centric biomedical associations, exposing hallucination sys

LLMsBiomedical NLPRAGHallucination

Post-Training, Not Data: Why LLMs Write the Way They Do

A critical analysis of how RLVR post-training shapes LLM linguistic patterns, and what AI detection systems reveal about automated assessment and human expressi

LLMPost-TrainingAI DetectionRLHF

Rotary GPU: Running 35B MoE Models on 8GB VRAM

A critical analysis of Rotary GPU, an exploratory approach to executing large Mixture-of-Experts models locally on consumer hardware with severely constrained V

LLM InferenceMixture of ExpertsEdge AIGPU Memory

Can LLM Reviews Be Gamed? Alignment and Vulnerability

New research tests LLM peer review alignment on 984 ACL submissions, finding limited reliability and that iterative LLM edits can game scores in specific scenar

Peer ReviewLarge Language ModelsNLPAcademic Integrity

Simplicial Message Passing as a Language Model Inductive Bias

The Cognitive Categorical Transformer reaches 21.27 PPL on WikiText-103, beating a matched GPT-2 Small baseline by 12% through category-theoretic structural pri

Language ModelsCategory TheoryArchitectureNLP

Why RL Preserves Circuits Better Than SFT During Fine-Tuning

New mechanistic interpretability research shows RL fine-tuning preserves internal attention circuits 13.5% better than SFT, explaining its resistance to catastr

Mechanistic InterpretabilityFine-TuningReinforcement LearningCatastrophic Forgetting

ICG: Personalised Cover Image Generation with MLLMs

ICG combines multimodal LLMs with diffusion models and multi-reward learning to generate personalised cover images without ground-truth labels.

Multimodal AIImage GenerationRecommender SystemsDiffusion Models

Modular LLM Architecture for Human Value Detection in Text

A three-module LLM pipeline detects and quantifies human values in text across any theoretical framework, showing architecture matters more than model choice.

AI EthicsNLPValue AlignmentLarge Language Models

Do LLMs Actually Introspect? A Critical Re-examination

New research challenges claims that LLMs can monitor their own internal states, showing that apparent introspection reduces to input-level pattern matching.

LLMsMetacognitionInterpretabilityAI Research

GEM: Hyperspherical Clustering for LLM Data Mixing

GEM reformulates LLM data curation as a variational problem on the hypersphere, improving downstream accuracy by up to 1.2% over strong baselines.

LLMData CurationPretrainingClustering

LLMs as Their Own Data Pipelines via Self-Verified Distillation

Self-Verified Distillation lets LLMs improve reasoning using only unlabeled prompts, achieving +16.7 points on math benchmarks with no external teachers.

LLMsSelf-TrainingReasoningSynthetic Data

LLMs Are Overconfident, Especially on Hard Tasks

A preregistered study of 11 LLMs finds systematic overconfidence that worsens with task difficulty, mirroring well-known human calibration biases.

LLMCalibrationReasoning ModelsAI Safety

Quantifying Redundancy in LLM Reasoning Chains

New research shows 61-93% of reasoning steps in frontier LLMs are redundant, and proves this overthinkin is a structural training artefact.

LLM ReasoningChain of ThoughtModel EfficiencyReinforcement Learning

Water-Filling Token Allocation for Reliable LLM Workflows

A new framework derives optimal token allocation policies for multi-agent LLM workflows, balancing latency, reliability, and cost using convex optimisation.

LLMMulti-Agent SystemsOptimisationAI Systems

FuRA: Full-Rank PEFT via Spectral Preconditioning

FuRA achieves full-rank weight updates with LoRA-level efficiency by freezing SVD bases, outperforming full fine-tuning on LLMs and VLMs.

Fine-TuningPEFTLLMsOptimization

LCF: Faster LLM Agent Communication via Latent Cache Transfer

Latent Cache Flow replaces text-based LLM agent communication with compressed KV cache transfer, cutting adapter size 73x and latency 8.5x over baselines.

LLM AgentsMulti-Model SystemsKV CacheInference Efficiency

The Positional Shortcut Hiding Inside Arithmetic CoT

Small language models copy the last number before an answer delimiter rather than reasoning through CoT steps. A new study exposes the mechanism across three ar

Chain-of-ThoughtMechanistic InterpretabilityArithmetic ReasoningLanguage Models

Bayesian GPs with Unknown Coordinates: Location Error Modelling

A technical deep-dive into Bayesian Gaussian process modelling where spatial coordinates are themselves uncertain, with applications to geostatistical inference

Bayesian InferenceGaussian ProcessesSpatial StatisticsGeostatistics

ICE's $25M Iris-Scanning Contract: Biometrics Without Oversight

ICE awarded Bi2 Technologies a $25.1M no-bid iris-scanning contract. A technical and policy analysis of the biometric system, its risks, and missing safeguards.

BiometricsSurveillanceAI PolicyCybersecurity

Feature Attribution Is Provably Unstable Under Collinearity

A new impossibility theorem proves no feature ranking can be faithful, stable, and complete when features are collinear, with formal Lean 4 verification.

Explainable AIFeature AttributionSHAPFormal Verification

Probabilistic Token Attribution for LLMs via Stochastic Processes

A new model-agnostic attribution score uses Bayes rule and stochastic process theory to explain LLM token importance without requiring model internals.

ExplainabilityLLMsNLPXAI

SOLAR: Autonomous Weight-Level Meta-Learning for Lifelong LLM Adaptation

SOLAR proposes a self-optimizing agent that treats LLM weights as an explorable environment, outperforming strong baselines across reasoning tasks without gradi

Continual LearningMeta-LearningLarge Language ModelsReinforcement Learning

Data Probes: A Principled Framework for Understanding LLMs

A position paper proposes synthetic data probes generated from known distributions to systematically study how data properties drive LLM behaviour and performan

Large Language ModelsData ScienceResearch MethodologyInformation Theory

How LLMs Misrepresent Disability: Toxic Positivity in AI

New research finds LLMs produce overly idealised portrayals of disability, erasing real struggles. A comparison with Reddit posts reveals systematic bias.

LLM BiasDisabilityAI FairnessNLP

MI-Guided Parallel Decoding in Masked Diffusion Models

A neural framework estimates pairwise mutual information from masked diffusion model hidden states, enabling 3-5x faster parallel decoding while preserving gene

Diffusion ModelsGenerative ModelsProtein DesignInference Efficiency

A 2D Framework for Classifying AI Agent Design Patterns

A new two-dimensional framework combines cognitive function and execution topology to classify 27 AI agent design patterns, resolving ambiguities in existing ap

AI AgentsLLMSystem DesignMulti-Agent Systems

AgentWall: Runtime Safety Enforcement for Local AI Agents

AgentWall intercepts AI agent actions before execution, enforcing declarative policies with 92.9% accuracy and sub-millisecond overhead across major coding envi

AI SafetyAutonomous AgentsSecurityDeveloper Tools

AI Annotator Safety Policy & Interpretability

How interpretability tools can resolve AI annotator disagreement in safety policies. A deep dive into arXiv research.

AI SafetyInterpretabilityMachine LearningData Annotation

Attention Maps Don't Predict VLM Reliability. Hidden States Do.

A mechanistic study of three VLM families finds attention sharpness is a near-zero predictor of correctness, while hidden-state geometry reaches AUROC>0.95.

Vision-Language ModelsMechanistic InterpretabilityAI ReliabilityAttention Mechanisms

Calibrated Uncertainty for Cost-Optimal LLM Cascade Routing

UCCI uses isotonic regression to calibrate LLM confidence scores, cutting inference costs 31% on a 75k-query NER workload while maintaining micro-F1 of 0.91.

LLMInferenceUncertainty CalibrationModel Routing

Differential Privacy and Social Bias in LLMs: A Complex Trade-off

New research finds DP training reduces bias in sentence scoring tasks but fails to generalise, revealing a gap between logit-level and output-level fairness.

Differential PrivacyFairnessLLMsNLP

GraphBit: Deterministic DAG Orchestration for LLM Agents

GraphBit replaces prompted LLM orchestration with a Rust-based DAG engine, achieving 67.6% on GAIA benchmarks with zero framework-induced hallucinations.

Agentic AILLM FrameworksMulti-Agent SystemsBenchmarks

HELLoRA: Smarter LoRA Fine-Tuning for MoE Models

HELLoRA targets LoRA adapters at the most-activated experts in MoE models, cutting trainable parameters by 84% while improving downstream task accuracy.

Fine-TuningMixture of ExpertsParameter EfficiencyLLMs

Hidden AI Coordinators Create Invisible Safety Risks in Multi-Agent Systems

New research shows invisible orchestrators in multi-agent LLM systems suppress protective behaviour and distort internal states, undetectable by output-based ev

AI SafetyMulti-Agent SystemsLLM ResearchEnterprise AI

LBW-Guard: Autonomous Governance Layer for Stable LLM Training

LBW-Guard adds a bounded control layer above AdamW to prevent training instability in LLMs, cutting perplexity 18.7% and surviving extreme learning-rate stress.

Machine LearningLLM TrainingOptimizationTraining Stability

Mistletoe: Attacking Speculative Decoding Without Changing Outputs

A new attack targets the acceptance mechanism of speculative decoding, collapsing inference speedup while preserving output quality and evading detection.

LLM SecuritySpeculative DecodingAdversarial MLInference Efficiency

OP-Mix: Unified Data Mixing Across the LLM Training Lifecycle

OP-Mix uses low-rank adapter interpolation to optimise data mixtures across pretraining and continual learning, cutting compute by up to 95% vs baselines.

Data MixingContinual LearningPretrainingLoRA

PREPING: Building Agent Memory Before Tasks Begin

PREPING tackles the cold-start problem in AI agents by constructing procedural memory from synthetic practice before any real tasks are observed.

Agent MemoryLLM AgentsAI PlanningSynthetic Data

Quantization Breaks Alignment: Bias Emerges in Compressed LLMs

New research shows 3-bit quantization causes 6-21% of previously unbiased LLM outputs to develop stereotypical bias, invisible to standard quality metrics.

LLMsQuantizationAI SafetyBias

Real-Time Diffusion Inference on Apple Silicon: What Works

A systematic 10-phase study reveals that CUDA optimisation insights fail on Apple M3 Ultra, achieving 22.7 FPS img2img via CoreML and model distillation.

Diffusion ModelsApple SiliconInference OptimizationCoreML

REVELIO: Uncovering Interpretable Failure Modes in VLMs

REVELIO systematically finds interpretable failure modes in Vision-Language Models using beam search and Gaussian-process sampling, revealing safety risks in au

VLM SafetyInterpretabilityAutonomous DrivingAI Research

Running Local LLMs on Apple M4 with 24GB RAM

The performance and architectural implications of running local LLMs on Apple M4 with 24GB unified memory. A technical deep-dive.

Local AIApple SiliconLLM InferenceEdge Computing

Scaling Laws for Skill Libraries in LLM Agent Systems

New research identifies logarithmic decay laws governing routing accuracy in LLM agent skill libraries, with actionable optimisations lifting pass rates signifi

LLM AgentsScaling LawsNLPAI Research

TeamTR: Trust-Region Fine-Tuning for Multi-Agent LLMs

TeamTR addresses a structural failure in multi-agent LLM fine-tuning, proving quadratic penalty scaling and achieving 7.1% gains over single-agent baselines.

Multi-Agent SystemsLLM Fine-TuningReinforcement LearningICML 2026

Trajectory-Balance Post-Training for Diffusion Language Models

TraFL addresses trajectory locking in diffusion LM post-training, improving mathematical reasoning and code generation across all benchmark settings.

Diffusion ModelsLanguage ModelsReinforcement LearningPost-Training

Variational Deep Embedding for Interpretable EEG Microstates

A new Conv-VaDE model jointly learns topographic reconstruction and soft clustering for EEG microstates, achieving 0.730 GEV with principled architecture search

EEGNeuroscienceDeep LearningClustering

Why LLMs Collapse to Narrow Outputs: A Calibration View

New research shows LLM diversity collapse stems from probability miscalibration during decoding, not just sampling heuristics. A framework across 14 models.

LLMsSamplingCalibrationNLP

Why Your Doctor Never Calls Back: AI in Healthcare Admin

How AI is tackling healthcare's hidden back-office crisis, and what it means for the future of medical admin and patient care.

AI in HealthcareMachine LearningHealthcare Automation

Cybersecurity (29 articles)

Russian Backdoor in Slovak Speed Cameras: A Supply Chain Case Study

Slovakia's NBU found SMS-triggered backdoors in Russian-made traffic cameras. A technical breakdown of the vulnerabilities and supply chain security implication

CybersecurityIoT SecuritySupply ChainCritical Infrastructure

Microsoft's 398-Vulnerability Patch Tuesday: AI's Double-Edged Sword

Analysis of Microsoft's August 2026 Patch Tuesday, covering 398 CVEs, the afd.sys zero-day, and the growing tension between AI-assisted vulnerability discovery

CybersecurityAI ResearchVulnerability ManagementWindows Security

Snowflake Extortions: Anatomy of a Cloud Credential Attack

Connor Riley Moucka's guilty plea exposes systemic failures in cloud authentication. A technical analysis of the Snowflake breach campaign and its security impl

CybersecurityCloud SecurityCredential AttacksData Breach

H96 TV Boxes: Ad Fraud, Proxy Networks & AI Deception

Bitsight's analysis of H96 streaming sticks reveals a sophisticated ad fraud operation using phone spoofing, AI-generated sites, and Blockly-automated bots.

CybersecurityIoT SecurityAd FraudAndroid TV

RFC 10015: Closing the Last Loopholes in TLS 1.2 Key Exchange

RFC 10015 formally deprecates RSA and finite-field Diffie-Hellman key exchanges in TLS 1.2, hardening the last remaining weak points before the protocol is reti

CryptographyTLSSecurity StandardsIETF

LG's Smart TV Proxy Ban: What the webOS Crisis Reveals

LG moves to ban residential proxy SDKs from its webOS app store after research found 42% of apps silently routing traffic through users' TVs. A technical deep-d

CybersecurityIoT SecuritySmart TVResidential Proxy

AI-Accelerated Vulnerability Discovery: The 570-Flaw Patch Problem

Microsoft's record 570-flaw Patch Tuesday exposes a critical tension: AI finds vulnerabilities faster, but also accelerates exploit development. An analysis of

CybersecurityAI ResearchMicrosoftPatch Tuesday

CISA GitHub Credential Leak: Six Months of Preventable Exposure

A deep technical analysis of CISA's 2026 GitHub credential leak, covering secrets management failures, incident response gaps, and what security teams must lear

CybersecurityIncident ResponseSecrets ManagementGitHub

IRIS C2 and the Zero-Day Market's Credibility Problem

A convicted felon is running a cybersecurity startup buying zero-day exploits for millions. What does this reveal about the offensive security market's structur

CybersecurityZero-Day ExploitsOffensive SecurityFraud

FBI Seizes NetNut and Popa Botnet: What It Means

FBI seizure of NetNut and the Popa botnet exposes deep flaws in residential proxy ecosystems. Technical analysis of the takedown and its wider implications.

CybersecurityBotnetsFBIResidential Proxy

Email Validation Via Spam Delivery: A Study in Absurdity

An analysis of a bizarre email validation scheme that routes addresses through a spam network, examining the technical and ethical failures involved.

EmailSpamCybersecurityWeb Development

Scattered Spider Guilty Pleas: Anatomy of a Cybercrime Network

Two Scattered Spider members pleaded guilty over the TfL cyberattack. An in-depth analysis of their TTPs, the SIM-swapping infrastructure, and what this means f

CybersecurityRansomwareSocial EngineeringSIM Swapping

Why Vulnerability Reports May No Longer Deserve Special Status

LLMs have disrupted the economics of security research. This analysis examines what that means for coordinated disclosure, open source maintainers, and triage w

CybersecurityOpen SourceLLMsVulnerability Disclosure

Popa Botnet: How a Proxy SDK Became a 2M-Node Threat

Analysis of the Popa botnet's architecture, its links to NetNut and Alarum Technologies, and what it reveals about the residential proxy economy's systemic risk

CybersecurityBotnetsResidential ProxiesAndroid Security

June 2026 Patch Tuesday: AI-Driven Vulnerability Discovery at Scale

Microsoft's record-breaking June 2026 Patch Tuesday reveals how AI-assisted vulnerability discovery is reshaping the threat surface. A technical analysis of the

CybersecurityVulnerability ResearchAI SecurityMicrosoft

Netherlands Seizes 800 Servers in Stark Industries Takedown

Dutch authorities arrested two men and seized 800 servers linked to Stark Industries Solutions, exposing how sanctioned Russian cyber infrastructure evades EU e

CybersecurityRussian Cyber OperationsEU SanctionsDDoS Infrastructure

Who Runs The Gentlemen Ransomware? An OSINT Deep Dive

Krebs on Security traces The Gentlemen ransomware admin through OSINT pivots, linking forum handles to Alexander Yapaev of Izhevsk, Russia. A technical analysis

CybersecurityRansomwareOSINTThreat Intelligence

Meta's AI Support Bot Exploited for Instagram Account Takeover

Pro-Iranian hackers exploited Meta's AI support bot to hijack Instagram accounts, exposing critical flaws in AI-driven account recovery workflows.

AI SecurityMetaInstagramSocial Engineering

ChatGPT for Google Sheets: Prompt Injection Data Exfiltration

PromptArmor's disclosure reveals critical indirect prompt injection vulnerabilities in ChatGPT for Google Sheets, enabling cross-workbook data exfiltration and

AI SecurityPrompt InjectionLLM VulnerabilitiesGoogle Sheets

CISA AWS GovCloud Keys Leaked on GitHub: A Deep Failure

A CISA contractor exposed AWS GovCloud credentials, plaintext passwords, and internal secrets on a public GitHub repo. Analysis of what went wrong and why it ma

CybersecurityAWSGitHubGovernment Security

Kimwolf Botnet: IoT DDoS Infrastructure and the Dort Arrest

Technical analysis of the Kimwolf IoT botnet arrest, examining DDoS infrastructure, device exploitation patterns, and the operational security failures that led

CybersecurityIoT SecurityDDoSBotnets

CISA Credential Leak: Insider Threat Meets Systemic Failure

A CISA contractor published AWS GovCloud keys and agency secrets to public GitHub. Analysis of the technical failures, institutional collapse, and CI/CD pipelin

CybersecurityCISAGitHubCredential Exposure

Anti-DDoS Firms Enabling the Attacks They Sell Against

The Brazilian anti-DDoS firm botnet scandal, exploring insider threat models, DDoS-for-hire economics, and cybersecurity ethics.

DDoSBotnetCybersecurityNetwork Security

Canvas Breach Analysis: EdTech Security Deep Dive

The Canvas data extortion attack targeting 275M students, exploring EdTech security failures and systemic vulnerabilities in education p

CybersecurityEdTechData BreachEducation Technology

FreeBSD execve() Privilege Escalation: Technical Analysis

FreeBSD-SA-26:13, a local privilege escalation via execve(). Deep technical analysis of kernel exec path vulnerabilities.

FreeBSDKernel SecurityPrivilege EscalationCVE Analysis

Patch Tuesday April 2026: 167 CVEs Analysed

Microsoft's record 167-CVE Patch Tuesday, the SharePoint zero-day, BlueHammer, and what this means for enterprise security in 2026.

Patch TuesdayCybersecurityMicrosoftZero-Day

Project Glasswing: How AI Is Rewriting Patch Tuesday

AI-assisted vulnerability discovery is driving record patch volumes across Apple, Google, Microsoft, and Oracle. An analysis of what this means for software sec

AI SecurityVulnerability ResearchPatch TuesdayAnthropic

Russia's Router Token Harvesting: A Deep Dive

Russia's GRU router exploit campaign harvesting Microsoft Office tokens across 18,000+ networks, implications for zero-trust security.

CybersecurityState-Sponsored HackingMicrosoft OfficeRouter Security

Scattered Spider's Tylerb Guilty Plea: Full Analysis

The Scattered Spider guilty plea, examining SMS phishing TTPs, social engineering at scale, and what this means for enterprise security

Scattered SpiderCybercrimePhishingSocial Engineering

Software Engineering (1 articles)

ClojureScript Gets Async/Await: What It Means

ClojureScript's new async/await support, analysing what this means for functional programming, AI tooling, and modern web development.

ClojureScriptFunctional ProgrammingAsync AwaitJavaScript

Technology (7 articles)

Teacher Housing as Infrastructure: What School Districts Get Right

School districts building affordable housing to recruit teachers reveals a systems-level insight about labour markets, urban economics, and public sector incent

Urban EconomicsLabour MarketsPublic PolicyHousing

The has_not_been_viewed_much Field: Obscurity in Cultural APIs

Exploring what the Art Institute of Chicago's boolean obscurity field reveals about digital curation, long-tail data, and the architecture of cultural APIs.

Open DataCultural ComputingAPI DesignDigital Humanities

Reversing Brain Aging with a Nasal Spray: What the Science Says

Texas A&M researchers report nasal delivery of a compound reverses brain aging markers in mice. A critical analysis of the methodology, mechanisms, and open que

NeuroscienceAgingBiotechnologyDrug Delivery

CreativityBench: Can AI Really Think Creatively?

CreativityBench, a new benchmark testing whether AI agents can repurpose tools creatively through affordance-based reasoning.

Artificial IntelligenceMachine LearningLLM BenchmarksCreative AI

Grid2Poster: Visualising Electrical Infrastructure from OSM Data

A technical analysis of Grid2Poster, an open-source tool that generates print-ready transmission grid posters from OpenStreetMap data using GeoPandas and Matplo

Open SourceEnergy InfrastructureGeospatialOpenStreetMap

OpenAI's New Voice API Features: Technical Analysis

OpenAI's new voice intelligence API features and what they mean for developers, security, and the future of AI.

OpenAIVoice AIAPIMachine Learning

Your Brain Processes Language While Unconscious

New research showing unconscious language processing and what it means for AI, NLP, and machine learning architectures.

NeuroscienceArtificial IntelligenceNLPMachine Learning

Detailed Research Areas

• AI Research — Large language models, autonomous agents, AI safety
• Security Research — CVE analysis, penetration testing, AI-powered security