acceptodds
Withdrawn from ICLR 2027

Does Wider Search Help? Autonomous Research For Improving Vision-Language Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

LLM-based research agents can autonomously propose hypotheses, modify training code, execute experiments, and substantially improve machine-learning models. Because model improvement is iterative, autonomous research can be viewed as search over research states and interventions, which raises the question of how much search control is actually needed. We study this question in autonomous post-training of a vision-language model by comparing a greedy single-trajectory policy () with a score-based multi-trajectory policy (). The two policies use the same research agent, starting model, training setup, evaluation pipeline, and research memory, and differ only in how many research states remain active. Both improve the aggregate score of the starting model from 0.576 to about 0.70, a relative gain of about 22% on average. However, under a round-aligned comparison that favors the wider policy, which trains three times as many candidates per round, reaches the plateau in fewer rounds but not a higher one: its best run (0.703) falls below the median greedy run (0.707). Analysis of the research histories shows that the wider policy retains diverse research states and attempts a broader range of intervention families, yet discovers no more productive ones, and that the effect of the same intervention can reverse sign as the surrounding training state changes, by margins well above training-seed noise. Our findings suggest that the large gains of autonomous post-training do not require wider search, and that exploiting additional preserved states, rather than generating them, is the central limitation for wider search in autonomous model research.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Autonomous Research Agents, LLM Agents, Model Post-Training, Vision-Language Models

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_BmXROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper asks whether keeping more research trajectories active helps an LLM research agent improve a model. Claude Opus 4.6 serves as the planner for autonomous post-training of a VLM2Vec-Qwen2-VL-7B embedding model, scored on a 12-task MMEB-V2 subset. A greedy single-trajectory policy (K=1) is compared with a top-K multi-trajectory policy (K=3); everything else is fixed: agent, memory, training setup, and evaluator. With three runs per policy over 22 rounds, both policies raise the score from 0.576 to about 0.70. K=3 plateaus sooner but not higher; its runs are more consistent with each other, and they explore more intervention families without finding more productive ones. A case study shows the same intervention (false-negative filtering) changing sign across parent states. The authors conclude that the bottleneck for wider search is exploiting the preserved states, not generating them.

Critical strengths

Clean controlled comparison: Only trajectory width varies; the agent, persistent memory, environment, and evaluator are held fixed. This isolates the search-control question better than most autonomous-research comparisons.

Useful process analysis: Distinguishing attempted breadth (coverage, entropy) from productive breadth is a helpful lens. It shows that wider search really does explore more, rather than just repeating the greedy policy.

Transparent reporting: Per-run numbers, discovered recipes, artifact construction, and explicit caveats about causal interpretation make the study auditable.

Critical weaknesses

The null result is underpowered: With three runs per policy, the mean difference (0.7054 vs 0.7016) is far smaller than the K=1 standard deviation (0.018). The study cannot distinguish “wider search doesn’t help” from a modest effect in either direction, and its headline “best K=3 below median K=1” rests on a single run.

Selection and reporting use the same benchmark: Every candidate, soup, and ensemble is selected on the same 12 tasks that produce the reported scores, and no held-out tasks are used. The 22% gain may therefore partly reflect selection on the benchmark, and K=3 performs three times more selections, which confounds the comparison.

Artifacts and single models are mixed in the comparison: The top K=1 run (0.722) is a 4-member soup and one K=3 run is a 5-model embedding ensemble. On single models alone the policies are 0.7015 vs 0.6965, which changes the framing, and the ensembles have different inference costs.

Weak support for the state-dependence claim: The evidence is one intervention with deltas of 0.002–0.006 from single runs. The “<0.001 seed noise” baseline comes from an unspecified number of seeds of one configuration, so it may not hold across states.

Narrow scope and limited novelty: The study covers one model, one benchmark, one planner, K∈{1,3}, and 22 rounds, with no compute-matched comparison (e.g., K=1 for 66 rounds). Toledo et al. (2025) already compared greedy and wider search strategies, so the generality and novelty of the conclusion are limited.

Questions for authors

Can you report how the final models from both policies perform on held-out MMEB-V2 tasks that were never used for selection?

How does the comparison look when restricted to single trained models, and with more runs per policy, or at least bootstrap confidence intervals?

How was the “<0.001 seed noise” estimated (how many seeds, which configurations), and does it hold for parent states other than the starting configuration?

Who assigned interventions to taxonomy families (authors or an LLM), and was this done blind to policy and outcome?

Reviewer_hq51OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper investigates using the novel tool of autonomous research for architectural search of vision-language models. The paper conducts a comparison between two methods of searching with the same underlying agent.

Critical strengths

Architectural search using autonomous agents is an interesting area of research given the recent rise in capabilities of agents. Methods to improve insights in this area are very important.

Critical weaknesses

One of my main concerns is that the main result in Figure 2 is a small-scale experiment. I would like significantly more results justifying the contributions. In figure 2, it seems like the variability from a single run (the low K=1 run) is pushing the early mean of K=1 down which is helping provide the headline result that K=3 reaches a plateau faster. A single run influencing a headline result indicates the need for significantly more evaluation. Additionally, a single K=1 achieves a much higher score than any K=3 run (noting that there are only three K=1 runs in total). This level of noise indicates that a more comprehensive setup is needed. Additionally, the experiments focusing only on K=1,3 in this limited setup and evaluation makes it difficult to establish broader claims about “wider search”.

In Section 3.2, the multi-trajectory policy seems to consist of a relatively simple implementation of a multi-trajectory policy. It would follow that this multi-trajectory policy might still reach a similar outcome to a single trajectory policy in fewer steps simply due to the implementation. If we simply scale a single trajectory policy across multiple processes in parallel then aggregate every step with a TopK selection, then this multi-trajectory policy could escape from local minima (assuming only a few trajectories became trapped in them) through the TopK selection whereas a single-trajectory policy needs to keep suggesting code edits to fix its mistakes and escape those minima. This leads me to believe that the main result of multiple trajectories achieving a similar plateau is potentially due to the formulation in Section 3.2 and that a more collaborative multi-search approach could help achieve a higher plateau. A different designer could have designed the multi-trajectory policy differently so that it achieves a higher score than the single-trajectory policy. A negative claim about a method like wider search in autonomous research needs a substantial amount of evidence across numerous runs, setups, and environments.

An evaluation of previous auto research methods along with validating their incorporation of wider search is also lacking. Demonstrating that previous methods without any wide search perform as well as their normal method would give more justification to the contribution.

Questions for authors
  • How can the authors distinguish between wider search being ineffective at achieving a higher score and the particular research agent being unable to exploit it under the specific task and specific formulation of wider search?
  • What would the results be if we significantly scale up evaluation to multiple different agents across many more runs under different formulations of wider search. For me, this would help support the claim about wider search in architectural search but would still require careful formulation and rigorous testing.
Reviewer_NV87OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether wider search improves the effectiveness of LLM-based autonomous research for vision-language model (VLM) post-training. The authors employ Claude Opus 4.6 as a research agent to autonomously improve VLM2Vec-Qwen2-VL-7B through iterative experimentation. Specifically, the paper compares greedy single-trajectory search (K=1) with multi-trajectory search (K=3), while keeping the research agent, training environment, and evaluation pipeline fixed. Experiments on 12 MMEB-V2 tasks show that both strategies achieve approximately 22% relative improvement over the initial model. Although K=3 explores more diverse research directions and exhibits lower variance across runs, it does not achieve better final performance than K=1. The authors conclude that effectively exploiting existing research states may be more important than simply increasing search width.

Critical strengths

Interesting Research Question. The paper investigates an important question in autonomous research: whether increasing search width leads to better model performance. Understanding the effectiveness of different search strategies is relevant to the development of autonomous research agents. Analysis of Research Behaviors. Beyond final performance, the paper examines exploration diversity, trajectory depth, productive intervention families, and state-dependent intervention effects, providing additional insights into the autonomous research process.

Critical weaknesses

Limited Novelty. The main contribution is an empirical comparison between K=1 and K=3 search strategies, without introducing a substantially new search algorithm or autonomous research framework. Previous work, particularly Toledo et al. [1], has already systematically investigated different search strategies for autonomous ML research agents and observed that more sophisticated search policies do not necessarily lead to better performance. The paper does not sufficiently clarify its novel contributions beyond these existing findings. Insufficient Experimental Comparisons. The paper only compares two relatively simple search configurations, K=1 and K=3, without benchmarking against more advanced autonomous research approaches, such as those explored in [1, 2, 3]. In particular, existing methods have investigated MCTS, evolutionary search, reflective memory, and targeted refinement. Although these approaches were primarily evaluated in different experimental settings, adapting representative methods to the same VLM post-training environment would provide a more convincing evaluation. Furthermore, the paper does not investigate other search widths or provide comparisons under equal computational budgets. Lack of In-Depth Analysis and Ablation Studies. Although the paper reports several research-process statistics, the analysis remains largely descriptive. For example, the authors attribute the limited effectiveness of wider search to difficulties in selecting appropriate state-dependent interventions, but provide limited experimental evidence to establish this explanation. The state-dependent intervention analysis is primarily supported by a small number of illustrative cases rather than systematic controlled experiments. Moreover, the paper lacks detailed ablation studies examining how research memory, intervention selection, and trajectory retention affect search performance. Limited Experimental Scope and Statistical Evidence. The experiments are conducted on a single VLM backbone using a subset of 12 evaluation tasks, with only three independent runs per search policy. This makes it difficult to assess whether the conclusions generalize to different models, tasks, or research environments. In addition, K=3 evaluates approximately three times as many candidates per round as K=1, making it important to distinguish the effects of search width, computational budget, and sequential feedback opportunities. Overall, while the research question is interesting, I believe the limited methodological novelty, insufficient experimental comparisons, and lack of rigorous analysis fall well short of the standards expected for acceptance at ICLR. References

[1] Toledo, Edan, et al. "AI Research Agents for Machine Learning: Search, Exploration, and Generalization in MLE-bench." Advances in Neural Information Processing Systems 38 (2025): 35309–35348.

[2] Chen, Jiefeng, et al. "MARS: Modular Agent with Reflective Search for Automated AI Research." Forty-third International Conference on Machine Learning, 2026.

[3] Nam, Jaehyun, et al. "MLE-STAR: Machine Learning Engineering Agent via Search and Targeted Refinement." Advances in Neural Information Processing Systems 38 (2025): 116692–116712.

Questions for authors

Novelty: How does this work fundamentally differ from existing studies of search strategies in autonomous ML research, particularly Toledo et al. [1], which already investigates Greedy, MCTS, and Evolutionary Search? Baseline Comparisons: Why are the experiments restricted to K=1 and K=3? Could the authors include more sophisticated search strategies, such as MCTS-based methods [1, 2] or targeted refinement approaches [3], adapted to the same experimental setting? Search Width and Computational Budget: How do the results change when comparing different search widths (e.g., K=2, K=3, K=5) under the same total number of training experiments? Can the authors disentangle the effects of search width and computational budget? Mechanistic Analysis: The authors suggest that wider search fails because the agent cannot effectively exploit alternative research states. What experimental evidence directly supports this explanation? Could the authors provide controlled ablation studies on state-conditioned intervention selection, research memory, and trajectory retention? Generalizability: Do the findings hold for other VLM backbones, benchmarks, or autonomous research agents? How robust are the conclusions given that each policy is evaluated with only three independent runs?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.