James Trappett

Paper Reviews, Security News & Technical Analysis

Latest Articles

Building Culturally Specific Stereotype Datasets with LLMs

12 July 2026

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

12 July 2026

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-PeerQUICAI Infrastructure

From Solvers to Research Agents: LLMs in Formal Mathematics

11 July 2026

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 VerificationTheorem Proving

ReCoLoRA: Spectral Consolidation for Continual LLM Fine-Tuning

11 July 2026

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

Continual LearningPEFTLoRALLM Fine-TuningCatastrophic Forgetting

Self-Distillation for Web Search Agents Without Teacher Models

11 July 2026

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-ImprovementNLP

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

10 July 2026

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-AGIReasoningAgentsBenchmarksLLMs

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

10 July 2026

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

EfficiencyTransformersMixture of ExpertsInferenceConditional Computation

When Does In-Context Search Actually Help LLMs?

10 July 2026

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

LLMsReasoningTheoryInferenceRLVR

Benchmarking KV-Cache Optimisations for Long-Context LLM Serving

9 July 2026

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 EfficiencyNLP

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

9 July 2026

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 AIStory Generation

FLORA: Fixing the Generator-Validator Gap in LLMs

8 July 2026

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 PaperLanguage Models

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

8 July 2026

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

LLMMultimodalOpen SourceEfficiencyReasoning

View all 165 articles →