- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This work focuses on teaching agents how to investigate a question and provide accurate, complete explanations. The authors note that successful trajectories may contain redundant or unproductive interactions, while rewarding final conclusions alone may overlook necessary experimental conditions. Thus, they propose GraphRover, a two-stage post-training framework. The first stage models dependencies within trajectories as an operation–evidence graph. It uses this graph to mask imitation loss for redundant or unproductive interactions while preserving the full interaction history. The second stage uses fine-grained rubrics for each question to assess how well answers satisfy requirements concerning findings, experimental conditions, and conclusions. The reinforcement learning reward combines item-level answer quality with the coverage of adequately satisfied requirements. Training with this framework substantially improves the performance of Qwen3.5-4B.
Critical strengths
- Training with the GraphRover framework leads to substantial performance gains for Qwen3.5-4B.
- The ablation studies provide strong support for the proposed mechanisms through controlled comparisons of dependency guidance, context preservation, and requirement coverage (Table 2). GGID outperforms both mask-budget-matched No-Graph Masking and Graph Pruning, while adding requirement coverage improves performance over optimizing mean rubric-wise answer quality alone on all three ablation benchmarks.
Critical weaknesses
- The graph represents dependencies among separate nodes for reasoning, tool calls, and observations, whereas each interaction receives a single functional-role label and confidence score . However, in steps 3 and 4 of Stage 1 in Figure 2, the REDUNDANT label appears to apply to individual nodes rather than entire interactions, suggesting an inconsistency.
- The abstract and introduction do not sufficiently articulate the limitations of closely related work. Although the paper cites several related studies, it provides limited analysis of their specific shortcomings and how these motivate the proposed approach.
- The experiments lack direct comparisons with closely related methods and omit relevant contemporary agents of comparable size, such as LiteResearcher-4B [1]. LiteResearcher-4B reports 78.0% on xbench-DeepSearch-2505, whereas GraphRover reports 75.0% on an unspecified XBench version, highlighting the need for comparisons under matched evaluation settings.
[1] Qu, B., Li, W., Pan, B., Zhang, J., Liu, Z., Zhang, P., Chen, W., & Zhang, B. (2026). LiteResearcher: A scalable agentic RL training framework for deep research agent. arXiv. https://doi.org/10.48550/arXiv.2604.17931
Questions for authors
- Appendix A.7 mentions checkpoint selection using a “development suite,” but does not define this suite or its relationship to the final evaluation sets. The default-setting results in Figure 3 exactly match the GGID results in Table 2, whereas the selected checkpoint’s FS-O score in Figure 5 (67.33) differs from that in Tables 1 and 2 (69.00). Please specify the data splits used for these figures and tables, clarify whether hyperparameter and checkpoint selection were independent of the final test sets, and explain the FS-O difference.
- The graph represents dependencies among separate nodes for reasoning, tool calls, and observations, whereas each interaction receives a single functional-role label and confidence score . The resulting supervision weight is shared by all assistant-generated tokens within that interaction. Could you clarify how and are determined when different components of the same interaction serve different functions—for example, when the reasoning is useful but the tool call is redundant?
Thank the authors for the detailed and transparent response, including the disclosure of the teacher and judge models and the additional final-checkpoint evaluation. After reviewing the new results, two concerns remain central to my assessment.
The CARL improvement depends on checkpoint selection and is indistinguishable from evaluation noise. The authors state that the 128-update checkpoint was selected based on its performance on FS-Research and FS-Olympiad, the same benchmarks reported in Tables 1–2, without an independent development set. The checkpoint trajectory is also non-monotonic. At 64 updates, CARL is below its GGID initialization on both FS-R (15 vs. 17 correct answers) and FS-O. It peaks at 128 updates and falls below GGID again at 256 (13 correct answers on FS-R; 58.00 vs. 63.00 on FS-O). The training run uses a learning rate of 2e-8, a single seed, and 60 / 100 test questions, and the reported checkpoint is the best of three. Under these conditions, the 128-update result is difficult to distinguish from noise. Under a consistent final-checkpoint rule, CARL improves over neither GGID nor CARL w/o Coverage, so the contributions of CARL and of the coverage term are not supported.
Most of the reported gains are attributable to distillation from Kimi 2.6. Plain SFT on the teacher trajectories already accounts for most of the improvement over the base model: +13 of +16 questions on FS-R, +29 of +32 on FS-O, and +8.02 of +11.33 F1 on DSQA. It also already exceeds the originally reported 35B references on all three benchmarks (25.00 / 66.00 / 71.43). Beyond SFT, GGID yields mixed changes (+2 and −3 questions on FS-R and FS-O, and +2.27 on DSQA). The final CARL checkpoint is below SFT on FS-R (−2 questions) and FS-O (−8 questions). HLE and XBench are reported only for the selected CARL checkpoint and not for SFT or GGID, so the contribution of the proposed components on these benchmarks cannot be assessed.
To support the claims, I would need: results on all five benchmarks for SFT, GGID, and each CARL checkpoint; denser checkpoints and multiple seeds, with selection on a held-out development set; and a comparison against compact baselines distilled from the same teacher. Without this evidence, I do not find that the proposed components provide a consistent gain beyond distillation, and I will maintain my score.