acceptodds
Under review as a conference paper at ICLR 2027

RubricAtlas: Decoupling Rubric Discovery from Fact Validation to Optimize Generalist Deep Research Agents

Abstract

Deep research agents are increasingly deployed to solve complex tasks by dynamically searching the web, integrating multi-source evidence, and composing long-form reports. Since such reports lack unique reference answers, task-specific rubrics are crucial for steering agent capability enhancement. However, existing rubric synthesis paradigms—such as offline model comparisons or online rollout evaluations—accept newly generated criteria without independent verification. This induces severe reward hacking, encouraging policies to favor excessive verbosity, superficial reasoning, and factually ungrounded demands, while existing datasets and queries suffer from limited domain coverage. To overcome these limitations, we present **RubricAtlas**, a unified framework for optimizing deep research agents across diverse knowledge domains. RubricAtlas builds a four-level taxonomy with 234 leaf concepts for diverse task synthesis. To enforce reliable supervision, we introduce Decoupled Rubric Discovery and Validation (DRDV), which decouples depth-oriented criteria discovery from search-augmented factual verification to filter ungrounded constraints. A Memo Context Manager maintains structured findings and evidence provenance throughout training and inference, supporting context integrity over long research horizons. For policy optimization, Group reward-Decoupled Normalization Policy Optimization (GDPO) normalizes rubric-wise advantages before aggregation, preserving fine-grained reward distinctions across varying reward scales. Extensive experiments show that RubricAtlas, with Qwen3-14B as its backbone, matches or surpasses strong proprietary models across three evaluated deep-research benchmarks, achieving an **82.60%** win rate on DeepConsult, a **46.67** RACE score on DeepResearch Bench, and **62.15%** rubric compliance on ResearchRubrics.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Deep Research Agent, Rubric Synthesis, Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_YV2hOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper presents RubricAtlas, a framework for training generalist deep-research agents. Its main components are taxonomy-guided task synthesis, Decoupled Rubric Discovery and Validation (DRDV), a Memo Context Manager, and rubric-wise reward normalization through GDPO. The authors construct 8,000 query-rubric pairs and train a Qwen3-14B model using cold-start SFT followed by RL. The resulting model performs strongly on DeepResearch Bench, DeepConsult, and ResearchRubrics, reaching a 46.67 RACE score, an 82.60% DeepConsult win rate, and 62.15% rubric compliance.

Critical strengths
  • Separating rubric discovery from factual verification is well motivated: criteria inferred from model outputs can otherwise reward verbosity or unsupported claims. Combining DRDV with a trajectory-grounded evidence reward gives the framework a coherent design beyond simply optimizing report style.
  • RubricAtlas-RL-14B consistently improves over its Qwen3-14B backbone and outperforms substantially larger open-source models across all three reported benchmarks. The component ablations also provide useful evidence that rubric validation, memory management, and criterion-wise normalization each contribute to the final performance. RL is effective.
Critical weaknesses
  • All three benchmarks focus mainly on long-form report generation and rely heavily on model-based judgments. It remains unclear whether the method also improves factual retrieval, search planning, and answer correctness on QA-oriented deep-search tasks such as BrowseComp, GAIA, or DeepSearchQA.
  • The paper reports strong downstream curves, but provides little direct evidence about the quality of the generated rubrics—for example, human audit results, validator agreement, difficulty. Moreover, the DRDV variants contain different numbers of rubrics, making it difficult to separate the effect of factual validation from rubric quantity or density.
Questions for authors
  • Can the authors report results on at least one fact-centric deep-search benchmark, such as BrowseComp, GAIA, or DeepSearchQA (if expensive, subset is ok)? This would help determine whether the gains extend beyond long-form rubric compliance to retrieval and reasoning accuracy.
  • Can the authors provide a human audit of the constructed rubrics, including factual correctness, relevance, atomicity, and agreement with the automatic validators? A rubric-count-matched comparison between unvalidated discovery and DRDV would also make the claimed mechanism substantially more convincing.
Reviewer_4j9DOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper explores the design of a rubric synthesis paradigm for deep research agents. The authors emphasize the importance of verifying the generated rubrics, as this can lead to reward hacking. As a response, the authors design RubricAtlas as a framework to synthesize tasks with a four-level taxonomy with 345 concept nodes and 234 leaf concepts. The authors then leverage DRDV and GDPO to demonstrate the effectiveness of the framework with an end-to-end pipeline on Qwen3-14B. The authors compare the trained model with some open/close-source baselines on DeepConsult, DeepResearchBench , and ResearchRubrics and show consistent gains. Besides, the authors also introduce Memo Context Management as a test-time tool that can improve both open-source models and the trained RubricAtlas-RL-14B.

Critical strengths
  1. This paper is well-written and easy to follow. The visualizations are high quality.

  2. The authors explore an end-to-end training pipeline that connects the core taxonomy-guided synthesis idea with data construction and training (SFT+RL). Experiments show consistent gain from metrics and rewards without high increase on report length.

  3. Rubric-based RL is an important topic for improving model capability on open-ended generation. The proposed reward design is nicely motivated and grounded (rubric verification in Section 3.1 and evidence reward in Section 3.3). Evidence from experiments in Section 4.5 also validates the effectiveness of trajectory-grounded reward.

Critical weaknesses
  1. One key problem of the setting is that the data construction essentially uses frontier models as the rubric generator or validator, which makes it hard to ablate whether the gains are from the pipeline or from distilling the strong teacher. It can be interesting to show in the analysis (1) if it is possible to set up the pipeline with the weaker teachers or the base model (Qwen3-14B) that will be trained on; (2) the same teacher with a budget-controlled simpler data pipeline.

  2. The authors include both proprietary models and open-source models in the experiments; however, the baseline comparison and ablations can be strengthened with (1) newer/stronger open/close-source models, e.g., GPT-6-Astra, Gemini 3.1 Pro/4 Argon, Opus 5.5, etc.; (2) full pipeline on other model sizes, e.g., 4B or 32B Qwen models.

  3. One main concern is that the connection between some components and the core rubric-construction idea is not sufficiently clear. For example, the Memo Context Manager appears to be a general context-management mechanism. Also, Section 3.3.2 devotes substantial space to the GDPO formulation from the previous work. Although these components can improve the overall system, the paper could benefit from (1) finding the connection between Memo and RubricAtlas; (2) providing more research insights through analyses directly related to its core design. For example, it would be interesting to examine (1) which types of rubric errors are corrected by validation and which corrections account for the gains; (2) how taxonomy-guided task and rubric construction interacts with the Memo Context Manager, and whether their combination provides additional benefits beyond the individual components; and (3) whether iterative multi-validator debate provides additional benefits over one-pass validation. Such analyses would help clarify what is learned from the proposed framework beyond the effectiveness of the loosely combined system.

Reviewer_b7qvOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes an end-to-end rubric generator that synthesizes evaluation criteria from multiple knowledge domains. By drawing on diverse domains, it aims to make rubric generation more independent of task-dependent synthesis pipelines. The paper is motivated by the potential for existing rubric-generation approaches to induce reward hacking, and evaluates the proposed method across different models and three public benchmarks.

Critical strengths
  1. The combination of broad task construction and evidence-checked rubric discovery addresses a meaningful weakness of training on open-ended research reports. Releasing a corpus of 8,000 query-rubric pairs could be useful independently of the final policy. The distinction between discovering discriminative criteria and checking their factual premises is well motivated.
  2. The empirical evaluation supports a useful end-to-end system. The 14B RL checkpoint improves RACE from 35.70 to 46.67 and rubric compliance from 39.21% to 62.15% (Table 1), while the discovery/validation, evidence-reward, GDPO, and memo ablations go beyond a single aggregate benchmark comparison. Appendix C also specifies distinct evaluation judges and three independent evaluation runs.
Critical weaknesses
  1. Direct evidence of reduced reward hacking is missing. The abstract motivates the approach by arguing that existing rubric-generation methods may lead to reward hacking, but I did not find experiments directly comparing the proposed method with existing approaches on this issue. The reported evaluation results do not directly establish whether this central problem is mitigated. This is the primary concern underlying my overall assessment.
  2. The validation ablation changes several consequential properties at once. Section 4.3 compares approximately 39 discovered criteria with approximately 31 validated criteria, while validation can also prune generic demands and recalibrate weights. With criterion-wise normalized advantages, removing or reweighting criteria changes the optimization signal even when their factual premises are unchanged. There is no count-matched pruning or non-factual-edit control to distinguish factual verification from reward simplification and verbosity filtering. This leaves the specific contribution of search-based fact validation less firmly established than the benefit of the complete filtering pipeline.
Questions for authors
  1. Request for core evidence: Could the authors provide a direct comparison of reward hacking between the proposed approach and existing rubric-generation methods, and explain how this phenomenon is defined, observed, and measured? On a blinded sample of accepted and rejected criteria, how often are factual premises unsupported, task-irrelevant, or incorrectly weighted before and after DRDV? If an independent human audit has already been conducted, please report its sampling procedure and results.
  2. What motivates the taxonomy of 13 roots, 345 concept nodes, and 234 leaf concepts? Beyond these counts, what evidence or coverage analysis supports the diversity relevant to the generated tasks and rubrics, and which domains remain poorly represented?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.