Parallel test-time scaling is one of the simplest ways to improve language model performance without any additional training: run multiple trajectories and aggregate. The assumption baked into this approach is that independent rollouts will explore meaningfully different reasoning paths. A new paper from Carnegie Mellon University and collaborators at Instituto Superior Técnico and NOVA School of Science and Technology challenges that assumption directly, showing that for agentic search systems, parallel threads often collapse onto near-identical first queries, retrieve overlapping evidence, and fail in correlated ways. The proposed fix, DivInit, is training-free, computationally cheap relative to the search budget, and consistently effective across a broad evaluation suite. The full paper is available at arXiv:2606.17209.
The Problem: Anchor Collapse in Parallel Search
Standard breadth scaling for search agents draws k independent trajectories under some sampling temperature and selects or aggregates their outputs. The implicit expectation is that stochastic sampling introduces enough variation to cover different retrieval paths. The paper identifies a failure mode they call anchor collapse: when a model is prompted with the same question k times, it tends to generate very similar first-turn queries. Because subsequent reasoning is conditioned on retrieved documents, near-identical initial queries produce near-identical evidence sets, and the trajectories diverge little from that point forward.
This is not merely a theoretical concern. The authors measure turn-1 query pairwise diversity (QPD) and find that standard parallel sampling concentrates near QPD values of approximately 0.2 on a Jaccard distance scale, while their method shifts the distribution toward 0.85. They also show that this first-turn separation propagates: threads that start with different queries maintain higher across-thread distance (ATD) across all subsequent turns, with a positive correlation between turn-1 QPD and ATD (Spearman rho of 0.42 for standard sampling). Increasing sampling temperature helps somewhat but saturates below the gains from explicit diversity selection, which rules out the simplest baseline response.
Methodology: DivInit
The intervention is deliberately minimal. Rather than drawing k first queries independently, DivInit draws a larger candidate pool of n queries from a single model call, then selects a subset of k < n using Maximal Marginal Relevance (MMR), a classic re-ranking criterion from information retrieval that balances relevance against redundancy. In the paper's configuration, n = 16 and k = 4, with the MMR trade-off parameter lambda set to zero, meaning pure diversity maximisation over the candidate pool. Each selected query then seeds an independent trajectory, and the remainder of the search loop is unchanged.
Several design choices are worth noting. First, generating 16 candidates in a single call is cheaper than running 16 full trajectories; the diversity selection happens before any retrieval occurs. Second, the authors deliberately restrict the intervention to turn 1 and show empirically that extending it to later turns yields negligible additional benefit, consistent with the anchor collapse framing. Third, MMR with lambda = 0 is a simple greedy selection procedure, which keeps the method easy to implement and interpret.
The agent architecture is a ReAct-style loop built on top of prior work, running for up to 8 turns per trajectory with BM25 retrieval over a local Wikipedia 2018 index for the multi-hop benchmarks, and live web search via SERPER for the open-web reasoning tasks.
Results
Evaluation covers five open-weight models (Qwen3 at 1.7B, 4B, and 8B; Gemma3 at 4B and 12B) across eight benchmarks. Key findings:
- DivInit improves over standard parallel sampling on nearly every model-benchmark combination at matched compute (same k, same token budget).
- Average gains on multi-hop QA benchmarks range from 5 to 7 percentage points, with the largest absolute improvements on Bamboogle and 2WikiMultihopQA.
- On open-web reasoning, WebWalker shows the strongest response to DivInit (6 to 7 points across all three evaluated models), while GAIA and HLE show smaller but consistent gains.
- Improvement scales with model capacity. For Qwen3, the average delta grows from 2.8 points at 1.7B to 7.4 points at 8B. The 1.7B model appears to hit a capacity floor where query diversification cannot be productively acted upon, a result that makes intuitive sense: if the model cannot reason differently given a different starting query, diversity at initialisation is wasted.
- Gains transfer to pass@1 aggregation under AggAgent, not just the pass@k ceiling metric, which is important for deployment relevance.
The temperature sweep experiment (Figure 3 in the paper) is particularly informative. Standard parallel sampling at tau = 2.0 still falls short of DivInit at tau = 1.0, which confirms that the mechanism is not simply about injecting more stochastic variation but about structured coverage of the query space.
Limitations and Open Questions
The authors are candid about several gaps. The pass@k metric measures the ceiling of the thread pool, not the accuracy of any single answer produced in deployment. While the AggAgent results are encouraging, the gap between pass@k and pass@1 remains substantial, and closing it is identified as an open problem. A thread pool with high structural diversity is only useful if there is an aggregation mechanism capable of identifying and combining the best evidence across threads.
The paper also restricts its analysis to search agents where the turn-1 action is a query string and the downstream effect is observable through retrieved documents. Whether anchor collapse manifests similarly in tool-use agents, code generation, or other multi-step settings is not addressed. Given that the mechanism is essentially about correlated conditioning on early actions, it seems plausible that analogous failure modes exist elsewhere, but that remains to be demonstrated.
There is also an implicit assumption that MMR over a single candidate pool is the right selection criterion. The lambda = 0 setting ignores relevance entirely in favour of diversity, which may not be optimal for all question types or retrieval settings. A more adaptive selection strategy that accounts for question difficulty or domain could potentially improve on the fixed configuration used here.
The authors gesture at a connection to reinforcement learning training: group-based methods like GRPO that rely on rollout diversity for learning signal may suffer from the same anchor collapse problem during training. If trajectories within a group are near-identical, the variance of rewards is low and gradient updates are weak. DivInit-style initialisation applied during RL training rollout collection is a natural extension worth pursuing.
Overall, this paper makes a clear and well-supported point: the standard assumption that independent sampling produces independent exploration is violated in agentic search, and a simple structured remedy exists. The intervention is practical, the evaluation is thorough, and the analysis of why it works is convincing. Anyone building or evaluating parallel search agents should account for query-level diversity at initialisation.