- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
In this paper, the authors propose ABSeeker, a credit assignment framework for training long-horizon search agents. The main motivation is to derive more fine-grained step-level signals for training frameworks to deliver signals on exactly which steps contributed to the successes/failures. To achieve that, the ABSeeker first performs Answer-backtracked clue recovery, which traces back from the answer to recover intermediate clues. Then, an LLM would score each step (reason + tool call step) with the recovered clues and assign a score to it. Based on the scores, the authors used them in two training paradigms: 1. ABC-SFT, where the scores are used as weights to losses. 2. ABC-GRPO, where scores are used as rewards. After training and applying context management, the scores improved upon the original 4B model to a range that authors claim to be on par with larger models around 30B.
Critical strengths
- This paper is well written and easy to follow. The main method is quite clear, and the illustrative figures aid in the understanding process.
- Step-level fine-grained signals is an important challenge to tackle in long-horizon tasks, not just for search agents. The authors propose a very interesting solution to it, which, from the experiments, looks like it works quite well.
Critical weaknesses
Despite its strong and clear motivation, I do have doubts about the experimental results:
- "Context Management" is not well-defined in this paper. It is a technique that increases the ABSeeker's performance a lot (by 10+%) so it is definitely very important. However, the authors only claims briefly in L407-410 that "we set the maximum context length to 256K tokens and apply the discard-all strategy for up to five rounds on BrowseComp and BrowseComp-ZH. " This claim reads like a test-time scaling hack to artificially boost up the numbers. Also, it looks like only MiroThinker and LongSeeker used this hack. Does this mean that all other models in Table 2 are reporting performances without using context management? That would make the results lose a lot of credibility and fairness.
- In Table 2, the authors should include Qwen-3.5-4B as a direct baseline as this was the backbone model the authors trained on. Without this baseline figure, it is hard to tell whether the agent actually gained performance improvement from the training or not.
- In the framework, an LLM is used for Answer-backtracked clue recovery and step scoring. This would create large computational costs for training. I don't see this cost analysis reported in the paper.
Questions for authors
Please see weaknesses
3 comments
Dear Reviewer 5eET,
Thank you again for your careful review and constructive feedback. As the discussion period approaches its end, we would like to briefly follow up on our rebuttal.
We have provided additional experiments and analyses addressing your three main concerns:
- Fair Comparison under Context Management (Q1): We clarified the discard-all strategy and the inference settings in Table 2. We also evaluated Standard GRPO and IGPO under matched context management settings and inference budgets, confirming that ABC-GRPO maintains its advantage beyond additional inference resources.
- Qwen3.5-4B Backbone Baseline (Q2): We added the original Qwen3.5-4B baseline across all five benchmarks, demonstrating consistent improvements from ABC training.
- Additional Computational Cost (Q3): We analyzed token usage over 1K paired samples, showing that the additional token overhead is 45.4% for ABC-SFT and 7.9% for ABC-GRPO, with clue recovery shared across eight RL rollouts.
We would appreciate your feedback on whether these additional results address your concerns and would be grateful if you could consider increasing your rating.
Thank you for your time and consideration.
The Authors
Follow-up on Rebuttal Responses and Reviewer Feedback
Dear Reviewers,
Thank you again for your valuable feedback. As the discussion period approaches its end, we would like to kindly follow up and ask whether our additional experiments and clarifications have addressed your concerns. We would greatly appreciate your feedback and hope you could consider increasing your ratings based on our additional results.
Thank you for your time and consideration.
The Authors
Thank you for your careful review and constructive feedback. We have conducted additional experiments and analyses to address your concerns, as detailed below.
1. [Q1] Fair Comparison under Context Management
Following MiroThinker and LongSeeker, we apply the discard-all strategy if the agent has not produced a final answer after 200 interaction turns, allowing it to continue searching with a fresh context window for up to five rounds, with a maximum context length of 256K tokens.
As stated in the caption of Table 2, *** denotes results without context management** and the corresponding results with context management are not officially reported. For a fairer comparison, we therefore also report our results without context management. To further address this concern, we additionally evaluated Standard GRPO and IGPO under the same context management strategy, 256K context window, and interaction budget as ABC-GRPO, using up to five rounds of discard-all.
Table 4. Performance comparison under matched context management settings.
All methods benefit from context management. However, ABC-GRPO maintains its advantage under matched inference budgets, outperforming IGPO by 3.9 points on BrowseComp and 2.4 points on BrowseComp-ZH.
These results show that ABC's improvements do not come only from additional inference resources. Its clue-anchored step-level supervision provides further gains under the same inference settings.
2. [Q2] Qwen3.5-4B Backbone Baseline
We evaluated the original Qwen3.5-4B under the same settings, without context management, and added its results as a direct baseline.
Table 3. Ablation study of ABC-SFT and ABC-GRPO (without context management).
The results show consistent improvements over the original backbone and confirm the effectiveness of our training framework.
3. [Q3] Additional Computational Cost
We analyzed the token usage of forward search, clue recovery, and step scoring over 1K paired samples.
Table 5. Token Cost Analysis of ABC (K Tokens).
Step scoring requires only 2.6% of the tokens used by forward search. Although clue recovery requires 42.9%, the recovered clues are shared across eight rollouts during RL, reducing the total additional token overhead to 7.9% for ABC-GRPO, compared with 45.4% for ABC-SFT.
Moreover, we use the relatively low-cost DeepSeek-V4-Flash for clue recovery and step scoring. The cost of tool API calls is negligible compared with LLM token costs in our setup. Considering the performance improvements, we believe this additional cost is reasonable and acceptable.
We appreciate your suggestions and hope these additional results address your concerns. Thank you again and we would be grateful if you could increase your rating.