acceptodds
Desk rejected at ICLR 2027

AREX: Verification-Guided Recursive Refinement for Deep Research

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Deep research requires agents to find answers that jointly satisfy multiple constraints. Discovering such answers is costly, whereas verifying a candidate can often be decomposed into tractable constraint-wise checks. This asymmetry suggests that agents can use the verification results of a provisional answer to direct subsequent research toward the constraints that remain unresolved. We introduce AREX, a verification-guided deep research agent that recursively refines its research state and answer. Its inner research loop gathers evidence, maintains structured records containing source URLs and supporting summaries, and returns a provisional answer, an evidence list, and an answer-level confidence score. The outer refinement loop uses this confidence to accept the result or, when confidence is insufficient, assesses the current trajectory and either launches a targeted refinement round or restarts the investigation. To support long-horizon research, AREX learns an autonomous context-update tool that consolidates its interaction history without relying on an external model. We train dense 4B and 122B-A10B MoE variants on verified synthetic tasks and high-quality trajectories, emphasizing decision-critical steps during both mid-training and reinforcement learning. Under a unified maximum inference budget, AREX achieves strong performance across six research and reasoning benchmarks, while additional recursive rounds yield consistent test-time scaling.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
deep research agent, Recursive Refinement

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_bSdxOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents AREX, a deep research agent that gathers evidence, assesses a provisional answer, and decides whether to accept it, conduct further research, or restart. The method combines recursive refinement with autonomous context updates and training that emphasizes important research decisions. Experiments on six benchmarks show strong performance at two model sizes, with further improvements from additional research rounds.

Critical strengths
  1. The approach addresses a practical problem in long research tasks: preserving useful findings while identifying what still needs to be investigated. The proposed research loop and context management are well aligned with this goal.

  2. The evaluation covers several types of search and reasoning tasks. The results at both model sizes provide evidence that the overall approach is useful.

  3. The scaling experiment supports the claim that the system benefits from additional research rounds, with accuracy increasing from 71.4% to 82.5%.

Critical weaknesses
  1. The comparisons with other systems need clearer reporting. The paper does not sufficiently explain where the baseline results come from or how their evaluation settings differ from AREX. This makes it difficult to determine how much of the reported advantage comes from the proposed approach versus differences in tools or inference resources. The internal comparisons provide stronger evidence for the method’s benefits than the cross-system comparisons.

  2. The key-step selection procedure needs a clearer description. Key-step supervision produces a substantial gain and is an important part of the training contribution. The paper describes the kinds of steps it selects, but how these steps are identified in practice remains unclear. This limits the reproducibility of an otherwise useful finding.

Questions for authors
  1. Which baseline results were reproduced, which were taken from published reports, and what important evaluation settings differ?

  2. How are key steps identified and labeled in the training trajectories?

1 comment

We thank the reviewer for recognizing the practical importance of our framework, the promising empirical results, and the effectiveness of our ablations. We address the concerns regarding evaluation transparency and reproducibility below.

W1/Q1. Baseline sources and evaluation settings.

All external baseline results in Table 1 are taken from publicly available technical reports and benchmark results, rather than reproduced in our evaluation environment. Only AREX-Turbo and AREX-Base are evaluated using our unified agent interface.

We have made efforts to align inference budgets with existing research agents. In particular, following MiroThinker MiroMindTeametal,˙2026MiroMind Team et al\., 2026, AREX adopts a maximum budget of 300 tool calls per round × 5 rounds = 1,500 tool calls Sec\. 5\.1; Appendix B.

While exact tool environments and computational costs cannot always be matched, particularly when external reports do not fully disclose their evaluation configurations, we align key inference-budget settings wherever possible. Table 1 therefore provides a broad performance reference rather than a strictly controlled comparison. The table also distinguishes full-HLE results from text-only results using the * notation.

Importantly, our conclusions regarding the effectiveness of AREX's components are primarily supported by controlled within-system ablations Figure2(bFigure 2(b and Table 2), where the relevant model and training or inference budgets are matched.

We have updated the Table 1 caption to explicitly clarify that all external baseline results are taken from publicly available technical reports or benchmark releases, rather than reproduced in our evaluation environment.

W2/Q2. Key-step identification and labeling.

As described in Appendix E, we use offline rule-based detectors to identify three types of decision-critical steps. The operational heuristics are summarized below:

  1. Evidence discovery: We select the first tool-calling step whose returned observation contains the reference answer after exploratory search. Matching uses normalized strings, ignoring capitalization and punctuation differences.

  2. Path rejection and redirection: We detect explicit rejection expressions eg˙,˙"ruledout,""doesnotfit"e\.g\., "ruled out," "does not fit" together with correction or alternative-candidate markers eg˙,˙"instead,""reconsider,""alternative"e\.g\., "instead," "reconsider," "alternative". We additionally require a search/visit action or adoption of the reference answer, with answer-bearing evidence already present in preceding tool observations.

  3. Key context updates: We select update_context calls that preserve a previously observed answer-related entity and at least one source URL, together with markers such as "verified," "evidence," "constraint," or "next step."

To reduce spurious annotations, these rules are applied only to trajectories with verified final answers and valid supporting tool observations Sec\. 3\.2; Appendix E. During training, the complete preceding context is retained, while the loss is applied only to the selected assistant step.

Our motivation for rule-based detection is practical: long research trajectories contain many routine interactions, making exhaustive manual step-level annotation prohibitively expensive. We therefore design conservative heuristics targeting observable events associated with evidence acquisition, search redirection, and context consolidation. These rules aim to identify a high-precision subset of informative decisions, rather than exhaustively determine the causal importance of every step.

Their training utility is supported by Table 2: replacing key-step supervision with equal-budget random-step replay reduces BrowseComp accuracy from 82.5% to 74.1% −84˙points-8\.4 points.

We will release the extraction code containing the complete matching patterns, filtering thresholds, and preprocessing details. As noted in our conclusion, developing more general and autonomous mechanisms for estimating step utility remains an important direction for future work.

We hope these clarifications address the reviewer's concerns.

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

AREX combines recursive research, autonomous context updating, and step-aware training for deep-research tasks.

Critical strengths
  1. The problem is important, and the framework is coherent.
  2. Results across six benchmarks and two model scales are promising.
  3. Existing ablations support the usefulness of context updating and the training components.
Critical weaknesses
  1. Unclear attribution of refinement gains. Increasing the number of rounds also increases the inference budget. Comparisons against independent restarts or confidence-based candidate selection under the same total budget are needed.
  2. Prior work already incorporates verification and learned context management. The current experiments do not clearly isolate the benefit of preserving research state and constructing targeted follow-up objectives.
Questions for authors
  1. Figure 2 shows improvements as the maximum number of research rounds increases, but the total tool-call budget also grows. Could the authors compare targeted refinement against independent restarts followed by confidence-based answer selection, using the same trained model, context-management mechanism, and total inference budget? This would help establish whether reusing and refining previous findings provides benefits beyond additional search.
  2. What specifically distinguishes AREX’s state-preserving, targeted refinement from existing verification-based agents such as MiroThinker-H1? Could the authors identify the main new mechanism and explain which experimental evidence isolates its contribution from the accompanying improvements in training and context management?
1 comment

We thank the reviewer for recognizing the importance of our work, the promising empirical results, and the effectiveness of our ablations. We address the concerns below.

W1/Q1. Recursive refinement versus additional inference compute.

We agree that increasing the number of research rounds also increases the available inference budget. Figure 2aa demonstrates test-time scaling 714˙71\.4% at K=1 to 82\.5% at K=5, but does not independently isolate the benefit of targeted refinement.

To address this concern, we conducted a controlled ablation, AREX w/o REFINE (Forced Restart). We disable the REFINE branch, forcing all low-confidence trajectories to restart from the original query. All other components remain unchanged, including the trained model, ACU, confidence-based acceptance and final answer selection, tool environment, and maximum inference budget (5 rounds × 300 tool calls).

For evaluation efficiency, we selected a 200-question difficulty-stratified subset of BrowseComp. Difficulty was determined by the number of correct outcomes in previous inference runs. We preserved the easy-to-hard ratio of the original 1,266-question test set, with questions randomly sampled within each difficulty group. Both methods were evaluated on the same subset.

Full AREX achieves 83.5% accuracy, compared with 79.5% for Forced Restart +40˙points+4\.0 points. This result supports the benefit of preserving and refining previous findings over independent restarts under the same maximum inference budget.

W2/Q2. Distinction from MiroThinker-H1 and mechanism attribution.

We acknowledge that MiroThinker-H1 already incorporates local and global verification and retains accumulated reasoning information within research episodes. Its global verifier can also request further reasoning when evidence is insufficient. However, its documented episode-restart mechanism reinitializes the agent from the original query, discarding the previous episode's state MiroMind Team et al\., 2026, Sec\. 3\.1, Eq\. 7.

AREX focuses on a complementary mechanism: state-preserving, targeted refinement across research rounds. For low-confidence results, the outer loop assesses trajectory recoverability and explicitly identifies reusable findings, unresolved issues, and a targeted objective for the next round. Recoverable trajectories are refined while preserving useful evidence, whereas unrecoverable trajectories trigger a clean restart Eqs\. 3–4; Appendix D\.3.

Thus, our contribution is not verification or context management in isolation, but using structured research-state assessment to guide transitions between research rounds.

Regarding attribution, Figure 2bb isolates the contribution of ACU under fixed round budgets, while Table 2 evaluates the training components under matched budgets. Our additional Forced Restart ablation further isolates the combined contribution of state preservation and targeted refinement while keeping the trained model and ACU fixed.

We hope these clarifications address the concerns regarding inference-budget attribution and better distinguish AREX from existing verification-based agents.

Reviewer_15yiOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a deep research agent which has an inner loop for research and an outer loop to refine the research or restart it based on model confidence. Authors also propose autonomous context updating, key-step-focused training, and step-aware RL.

Critical strengths

One strength of this paper is that improving long-horizon deep research agents is a practically useful problem. The empirical results are strong across many benchmarks, and for relatively small models. Ablations on context updating and training strategies are useful.

Critical weaknesses
  1. Arex does not feel novel to me since there has been extensive work on verification-guided search, self-refinement, and context management. For example, asymmetric verification (https://arxiv.org/pdf/2510.06135), ReSum and memory as action. It is unclear what new learning this paper brings to the community beyond combining several existing techniques
  2. The main motivation fo the paper is the asymmetry between verification and discovery. Hoever, is there an actual independent verifier in the proposed framework? The outer loop replies on model confidence which makes it similar to self-reflection or confidence-based retrying. How reliable is this signal for guiding refinement? Can the authors compare against an actual verifier or simpler self-critique baselines?
  3. It is unclear whether the improvement is due to the proposed recursive refinement or simply more test-time compute. Can the authors compare against a single longer research trajectory, or multiple independent research attempts with the same compute budget?
  4. The key-step-focused training results are interesting, but the authors use rule-based detectors to identify important steps, and it is unclear how accurate these detectors are. Higher loss on these steps does not necessarily mean they are more important for final task success. Also, the paper does not provide enough details on the amount of training data, teacher models, and training configurations. Additionally, is the comparison against other research agents done with matching inference time compute and matching environment?
Questions for authors

See weaknesses

2 comments

We thank the reviewer for recognizing the practical importance of our work, the strong empirical results, and the useful ablations. We address the concerns below.

W1. Novelty and distinction from prior work.

We agree that verification, self-refinement, and context management have important precedents. Asymmetric verification primarily investigates verifier-based test-time scaling and candidate selection; ReSum focuses on context summarization; and Memory as Action learns memory-editing operations.

AREX addresses a complementary problem: how to transform an incomplete research result into an actionable next research round. Specifically, its outer loop identifies reusable findings, unresolved issues, and a targeted follow-up objective, enabling the agent to ACCEPT, REFINE, or RESTART Eqs\. 2–4. Unlike simply extending or restarting research, REFINE preserves useful evidence while redirecting exploration toward remaining constraints.

Our contribution lies in this structured inter-round research control, together with autonomous context updating and step-aware training. The additional controlled experiments below further evaluate its benefits.

W2. Independent verification and confidence reliability.

The reviewer is correct that AREX does not employ a separately trained inference-time verifier. Instead, the research model produces an evidence-linked confidence score reflecting answer completeness, consistency, provenance, and temporal validity Sec\. 2\.2\.2.

Appendix H.1 (Figure 4) provides evidence that this signal is informative: with ACU, 95.9% of correct outputs receive confidence scores in [90, 100], compared with 33.0% of incorrect outputs. While this does not establish probabilistic calibration, it demonstrates useful discrimination between correct and incorrect answers.

We additionally evaluated a simple prompt-based self-critique baseline on the same 200-question subset described below. Full AREX achieves 83.5% accuracy, compared with 75.0% for self-critique. Our error analysis suggests that self-critique frequently fails to identify incorrect provisional answers.

We acknowledge that this baseline was not separately fine-tuned, and more sophisticated self-critique strategies may perform better. Our design instead adopts a lightweight approach: the research model directly produces confidence alongside its answer and supporting evidence, without requiring a separate critic model or an additional specialized verification pipeline.

W3. Recursive refinement versus additional test-time compute.

We agree that Figure 2aa alone does not distinguish recursive refinement from increased inference computation. We therefore evaluated several alternative inference strategies under the same maximum tool-call budget.

For evaluation efficiency, we used a 200-question difficulty-stratified BrowseComp subset. Difficulty was defined by the number of correct outcomes in previous inference runs. We preserved the easy-to-hard ratio of the original 1,266-question test set and randomly sampled within each difficulty group.

The results are:

Inference strategyAccuracy
Single extended research trajectory~75.0%
Additional independent restarts for unfinished trajectories77.0%
AREX w/o REFINE (Confidence-Guided Forced Restart)79.5%
Full AREX83.5%

For the extended-trajectory baseline, we allowed additional update_context opportunities for unfinished trajectories. However, qualitative inspection suggests that unsuccessful trajectories often remain trapped in the same incorrect research direction. Moreover, trajectories that have already finished cannot benefit from simply extending the allowed trajectory length.

Allowing additional independent restarts improves accuracy to 77.0%, but this strategy still primarily benefits unfinished trajectories.

In contrast, AREX separates the completion of an inner research round from termination of the overall research process. Low-confidence provisional answers can trigger further investigation, including cases where a trajectory has already finished but produced an incorrect answer. This helps explain the improvement to 79.5% with confidence-guided forced restarts.

Full AREX further improves accuracy to 83.5% by preserving reusable findings and constructing targeted objectives when trajectories are recoverable, rather than always restarting from scratch.

These results suggest that AREX benefits from adaptively deciding when to continue, refine, or restart research, rather than merely allocating additional computation to longer or independent trajectories.

W4. Key-step detection, training details, and evaluation fairness.

Key-step detection. Our detectors are intentionally heuristic rather than ground-truth annotations of causal step importance. They identify observable events such as answer-relevant evidence discovery, explicit candidate rejection and search redirection, and meaningful context updates (Appendix E). Labels are retained only for trajectories with verified final answers and valid supporting tool observations.

Because long research trajectories contain many interactions, exhaustive manual assessment of step importance is prohibitively expensive. Moreover, there is no canonical step-level ground truth for causal importance. We therefore use conservative, rule-based extraction to identify a subset of informative research decisions.

We agree that higher step loss alone does not establish causal importance. Our stronger evidence is the controlled ablation in Table 2: replacing key-step supervision with equal-budget random-step replay reduces BrowseComp accuracy from 82.5% to 74.1% −84˙points-8\.4 points. As noted in our conclusion, developing more dynamic and autonomous mechanisms for estimating step utility remains an important direction for future work.

Training reproducibility.

Training data and configurations. We have expanded Appendix B to provide additional training details. Agentic Mid-training uses over 10,000 complete SFT trajectories and over 10,000 key-step supervision instances, collected with the assistance of teacher models including GLM-5 and Kimi-K2.6. Unlike full-trajectory SFT, each key-step instance retains the preceding trajectory as context but applies the training loss only to the selected assistant step. Both use the same training hyperparameters. The RL stage uses over 1,000 training questions.

The updated Appendix B also reports detailed training configurations for both Agentic Mid-training and RL, including learning rates, batch sizes, training steps, and other relevant optimization settings.

Evaluation fairness. All external baseline results in Table 1 are taken from publicly available reports rather than reproduced within our environment. We have made efforts to align inference budgets with strong research agents, particularly MiroThinker, following a maximum budget of 300 tool calls per round × 5 rounds Sec\. 5\.1. While exact tool environments and computational costs may differ across systems, our conclusions about individual AREX components are primarily supported by controlled within-system comparisons.

We hope these clarifications and additional experiments address the reviewer's concerns.

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

Related papers

Discussion (0)

Sign in to comment.