acceptodds
Desk rejected at ICLR 2027

Iris: Climbing to the Search Frontier

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

Abstract

We present Iris-mini and Iris-pro, two search agents trained at the 35B-A3B and 397B-A17B scales, together with the data pipeline and training recipe behind them. Tasks are reverse-constructed from the hyperlink structure of a web corpus: we author multi-hop chains over an entity graph distilled from a seed page and its out-links, rewrite every non-answer entity into a descriptive reference, and admit only questions that a reference model fails closed-book yet solves once the supporting evidence is supplied. These questions are then turned into trajectories, which are filtered at the trajectory- and turn-level before supervised fine-tuning (SFT). The policy is then optimized by reinforcement learning (RL) against live search. We alternate the two stages in a procedure we call SFT-RL climbing, returning the hardest solved and most efficient rollouts of each RL round to the next supervised pass. Because inference-time context management is worth more on these benchmarks than most reported differences between systems, we evaluate every benchmark both with and without it, holding the tool set, the context limit, and the judge fixed. All results come from a single ReAct agent. With management enabled, on BrowseComp, BrowseComp-ZH, DeepSearchQA, and Humanity's Last Exam the two models reach and , the strongest overall results among open-source search agents in their respective parameter ranges. We plan to release the model weights together with the complete recipe for data construction, training, and evaluation.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Agentic Post-Training, Search Agents, SFT–RL Climbing

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_6HuzOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents Iris-mini and Iris-pro, two search agents trained at the 35B-A3B and 397B-A17B scales, respectively, using an end-to-end training pipeline for long-horizon information seeking. The proposed pipeline consists of three main components. First, the authors construct challenging multi-hop search tasks from web hyperlink structures by extracting entity graphs, generating reasoning chains, abstracting explicit entity references, and filtering questions based on difficulty and solvability. Second, teacher-generated search trajectories are refined through trajectory-level filtering and turn-level loss masking before supervised fine-tuning. Third, the agents undergo reinforcement learning with live web search, followed by iterative SFT–RL climbing, where successful and efficient trajectories from RL are incorporated into subsequent SFT rounds. The authors evaluate both models on BrowseComp, BrowseComp-ZH, DeepSearchQA, and Humanity's Last Exam. They additionally investigate the impact of inference-time context management, comparing settings without context management against discard-all, retry, and their combination. The proposed systems achieve strong performance across the four benchmarks, with particularly notable improvements on BrowseComp. The results also highlight the substantial influence of context management on the measured capabilities of long-horizon search agents.

Critical strengths
  1. Strong empirical performance. Iris-mini and Iris-pro achieve competitive results across four challenging search benchmarks, outperforming most comparable open-source agents at their respective model scales.
  2. Insightful context management analysis. The paper evaluates performance with and without CM, revealing substantial gains (e.g., +17.5 points on BrowseComp for Iris-mini) and highlighting the importance of inference-time strategies.
  3. Practical iterative training framework. The proposed SFT–RL climbing strategy incorporates successful RL trajectories into subsequent SFT rounds, offering an intuitive approach to progressively improving search capabilities.
Critical weaknesses
  1. Missing ablations for key components. No controlled experiments isolate the contributions of data synthesis, turn-level filtering, or iterative SFT–RL climbing. In particular, the advantage over standard SFT→RL remains unverified.
  2. Insufficient reproducibility details. Important information is missing, including training data sizes, teacher model identities, filtering thresholds, climbing iterations, and RL hyperparameters.
  3. Incomplete inference-cost analysis. CM provides substantial gains, but token consumption, tool-call counts, and compute-matched comparisons are absent. The role and cost of the auxiliary 397B summarizer during evaluation also remain unclear.
Questions for authors
  1. Effectiveness of iterative climbing. Could the authors report performance across existing SFT–RL climbing rounds or intermediate checkpoints? This would help clarify whether iterative climbing provides consistent improvements over the training process.
  2. Data quality and verification. How reliable is the proposed dual-criteria verification in ensuring genuinely challenging and unambiguous multi-hop questions? Have the authors conducted any quality assessment of the synthesized data?
  3. Training transparency. Could the authors provide more details on the training data scale, teacher models, number of climbing iterations, and trajectory selection statistics? These details would help clarify the practical training recipe without requiring additional experiments.
  4. Inference efficiency and fairness. Could the authors clarify the effective inference budgets across different CM strategies, including token consumption and tool calls?
3 comments

Thank you very much for your careful review. We greatly value your comments and suggestions.

W1 + Q1

We apologize that we cannot disclose all training details at this stage. However, we can provide a clearer account of the climbing schedule and the role of each stage.

We perform two SFT–RL climbing rounds: SFT → RL → SFT → RL. The first SFT stage builds the foundational capabilities needed for search-agent tasks and contributes the largest improvement in benchmark scores. The following RL stage further develops these capabilities through exploration. In the second SFT stage, we reinforce useful patterns found in successful RL rollouts. The immediate benefit is greater efficiency rather than higher benchmark scores. This improved efficiency then supports exploration of more challenging questions during the second RL stage, leading to further performance gains. Under evaluation without any context management, completing both climbing rounds consistently improves performance by more than 2% on BrowseComp, BrowseComp-ZH, DeepSearchQA, and HLE compared with stopping after the first SFT–RL round.

Q2

This is a particularly important question. We agree that assessing the quality of synthetic data is essential.

Regarding difficulty, when Kimi-K2.6 attempts questions generated by our pipeline, successful attempts require more than 50 tool-call turns on average, while unsuccessful attempts involve around 150 turns. These figures illustrate the substantial search effort involved (note that this does not mean the teacher model used in training is Kimi-K2.6).

Regarding ambiguity, we distinguish between incorrectly constructed questions and questions with multiple valid answers. Dual-criteria verification helps filter out incorrect questions, but it does not guarantee that every retained question has a unique answer. Therefore, before RL training, we generate three rollouts per question using a strong model and retain only questions for which at least two rollouts are judged correct and produce the same answer. This provides an additional check on answer consistency, although it is not a guarantee of uniqueness.

We will release further details about the synthetic data, the complete data construction pipeline, and a subset of the QA pairs and trajectories upon acceptance.

W2 + Q3

We fully understand your point that these details would help readers understand the training procedure without requiring additional experiments. This is a completely reasonable request, and we sincerely apologize that we cannot disclose all of this information at present.

We have clarified the number of climbing rounds above. Details concerning the training data and teacher models remain unavailable for disclosure at this stage, but we remain committed to releasing the complete data construction pipeline and training recipe.

W3 + Q4

Thank you for highlighting the need for precise budget accounting. Although detailed cost and configuration information is not available for every comparison model, we can clarify our own evaluation settings:

  • Under discard-all, the context-reset threshold is 128K tokens, consistent with DeepSeek-V3.2. The maximum cumulative budget is 600 tool-call turns per question. Turns consumed before a reset remain counted, and the attempt stops immediately when the total reaches 600.
  • Under retry, each new attempt appends a summary of the preceding failed attempt to the task prompt. The stopping rule is based on the number of attempts rather than a cumulative turn budget: we allow up to five ReAct attempts per question.
  • Under discard-all + retry, the inner loop allows up to 600 cumulative tool-call turns per attempt, while the outer loop allows up to five attempts. This configuration can consume substantially more inference resources, so we include it only as an exploratory analysis, rather than as the setting used for the main results table.

The 397B summarization model serves two main purposes: filtering irrelevant webpage content, such as advertisements, and summarizing retrieved pages to reduce the amount of context consumed.

Our evaluation primarily focuses on tool-call turns. We are also willing to release all evaluation trajectories to support a more detailed analysis of token consumption.

These settings describe our own evaluation protocol. Different teams use different resource allocations, API configurations, and inference budgets, and these details are not always fully disclosed. We therefore cannot claim that all comparisons are matched for total token consumption. However, we take responsibility for the accuracy of every result we report for our own models.

We will revise the manuscript to address your suggestions and notify you of the specific changes and their locations as soon as the revision is ready. We welcome any further questions or discussion.

Thank you for your rebuttal. After reading the response carefully, I think my questions remain partially resolved. Therefore, I will maintain my rating.

Dear Reviewer 6Huz:

Thank you very much for your follow-up and for taking the time to consider our responses. The revised manuscript has been submitted. Thank you again for your thoughtful feedback and consideration.

Best regards, The Authors

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

This paper aims to improve LLM search capabilities through synthetic multi-hop search tasks and an iterative SFT–RL training pipeline that incorporates successful RL trajectories into subsequent supervised training. The resulting Iris-mini and Iris-pro models achieve strong performance across four challenging search benchmarks.

Critical strengths

The paper presents strong search agents at both the 35B and 397B scales, with competitive results across four challenging benchmarks. These results demonstrate the practical value of the overall training pipeline for long-horizon search.

Critical weaknesses
  1. The pipeline largely combines established practices, but the paper does not sufficiently isolate their contributions. In particular, without a comparison between continued RL and SFT–RL climbing under matched training budgets, it is unclear how much additional value the proposed alternation provides.

  2. The quality assurance for synthetic questions is insufficiently established, particularly regarding answer uniqueness. A reference model producing the intended answer given the entity graph does not rule out other valid answers, especially if the graph omits relevant entities or relations. The lack of concrete examples and quality audits makes this concern difficult to assess.

Questions for authors

How is answer uniqueness verified beyond checking whether a reference model produces the intended answer given the entity graph? Could the authors provide representative examples and an audit of ambiguous or incorrect questions, including how cases with multiple valid answers are detected and filtered?

4 comments

This is an excellent question! We were even excited you raised it.

You are right. The entity-graph-based QA pipeline cannot guarantee answer uniqueness, no matter how careful we are. This is especially important for RL, where a valid alternative answer may receive no reward because it differs from the reference answer.

To reduce this risk as much as possible, we apply two stages of filtering:

  1. During QA synthesis, we ask an LLM with tool access and the full entity graph to answer each question. We discard questions for which it fails to recover the reference answer. This helps flag potentially flawed or ambiguous questions.
  2. Before RL training, we generate three rollouts per question using a strong model and retain only questions for which at least two rollouts are judged correct and produce the same answer. This provides an additional check on answer consistency. (Separately, we run eight rollouts per question using the initial policy model to select questions within a suitable pass-rate range. However, note that this step controls training difficulty and should be distinguished from the consistency filtering above.)

We will clarify these procedures and their limitations in the revised manuscript. We are quite excited about this issue, so we wanted to respond to you right away. We will follow up with the exact locations of the changes once the revision is ready.

Regarding W1, we fully appreciate your concern. Although we are not yet able to disclose all details, we can clarify that training consists of two SFT–RL climbing rounds: SFT → RL → SFT → RL.

Under evaluation without any context management, the initial SFT stage delivers the largest score gains, and the first RL stage further improves performance. The subsequent SFT stage reinforces good patterns discovered during RL rollouts, making the model more efficient, although this improvement does not immediately translate into higher benchmark scores. This greater efficiency enables the model to explore more challenging questions during the second RL stage, which yields further gains across the evaluated benchmarks.

We have released the model weights and the harness. Upon acceptance, we commit to releasing the complete data construction pipeline, the full training recipe, and a subset of the QA pairs and trajectories to support reproducibility and further research by the community.

Thank you for the clarifications; my concerns on W1 and W2 remain only partially addressed, so I will keep my rating.

Dear Reviewer RTxK:

Thank you very much for your follow-up and for taking the time to consider our responses. The revised manuscript has been submitted. Thank you again for your thoughtful feedback and consideration.

Best regards, The Authors

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

This paper introduces the Iris-mini and Iris-pro search agents, combining web-graph-based construction of multi-hop questions, trajectory and turn filtering, reinforcement learning with online search, and alternating SFT–RL training. Both model sizes perform strongly on four search and reasoning benchmarks, with results reported both with and without context management.

Critical strengths

1.Both model sizes achieve strong results. For example, Iris-pro reaches 88.6 on BrowseComp and 92.9 on DeepSearchQA, demonstrating the practical value of the overall training approach (Table 1).

2.Reporting performance separately with and without context management is valuable. Table 2 clearly shows that this mechanism has a large effect on long search tasks such as BrowseComp, helping explain the sources of the final system's performance.

Critical weaknesses

1.The experiments do not establish the independent contributions of the main training choices. The current analysis focuses on context management, without comparing SFT alone, SFT followed by a single round of RL, continued RL, and alternating training. The benefit of turn filtering is also not tested, leaving it unclear which design choices drive the final gains.

2.Evaluation budgets are insufficiently specified. Section 4.1 states that configurations are consistent, but does not provide the maximum number of turns, reset thresholds, or cumulative generation budget, nor does it clearly identify which baselines in Table 1 were rerun under the same framework. Resets or retries increase the actual amount of search, so fixing the context window alone is insufficient to assess whether the comparisons are fair.

3.The training recipe lacks key details, including the amounts of synthesized and retained data, the identities of the teacher and verification models, filtering thresholds, RL parameters, and the number of alternating rounds. Section 3.3 also defers these details to a future release. The current version does not provide enough information for a thorough assessment of the training method.

Questions for authors

1.With the same training data and compute budget, what gains does alternating SFT–RL provide over continued RL? Does turn filtering also make an independent contribution? 2.After a context reset, are previously consumed tokens and tool calls counted toward a single overall budget? Which main baselines in Table 1 were reevaluated with the same tools and total budget?

3 comments

Thank you very much for your careful review.

W1 + Q1

This is a very good question. We apologize that we cannot disclose all training details at this stage, but we can clarify the rationale behind alternating SFT–RL climbing and the gains we observed.

Our training consists of two SFT–RL climbing rounds: SFT → RL → SFT → RL. The initial SFT stage establishes the model’s broad capabilities for search-agent tasks and therefore delivers the largest score gains. The first RL stage further improves performance. The subsequent SFT stage reinforces good patterns discovered during RL rollouts, making the model more efficient, although this improvement does not immediately translate into higher benchmark scores. This greater efficiency enables the model to explore more challenging questions during the second RL stage. This is the rationale for alternating SFT and RL rather than continuing RL alone. Under evaluation without any context management, the model trained with two SFT–RL climbing rounds consistently outperforms the model trained with a single SFT–RL round by more than 2% on BrowseComp, BrowseComp-ZH, DeepSearchQA, and HLE.

Turn-level filtering also provides clear benefits. We conducted comparisons at three training-data scales and found that adding fine filtering consistently improves performance by 1–2% across search benchmarks compared with coarse filtering alone.

W2 + Q2

Consistent with the practice used by DeepSeek-V3.2 and other frontier models, tool-call turns remain counted toward the total budget after a context reset. We stop answering the question as soon as the cumulative turn limit is reached. We generally do not impose a separate cap on cumulative token consumption.

It is not feasible for us to re-evaluate every comparison model using exactly the same tools and total budget. First, not all model weights and evaluation harnesses are publicly available. Second, different teams use different search APIs, tool configurations, and evaluation settings, and not every detail is disclosed.

For our internal leaderboard, we evaluate models under a unified setup. For public comparisons on established benchmarks such as BrowseComp, however, we report the scores published by the respective teams. We cannot independently guarantee the evaluation details behind other teams’ results, but we take responsibility for the accuracy of our own reported numbers.

W2 + W3

We apologize that we cannot disclose the full training details at this stage. We have clarified the number of climbing rounds above, but information such as the training-data scale and teacher models is not yet available for disclosure. We have already released the model weights and the evaluation harness. Upon acceptance, we commit to releasing the complete data construction pipeline, the full training recipe, and a subset of the QA pairs and trajectories to support reproducibility and further research by the community.

We will revise the manuscript to address your questions and notify you of the specific changes and their locations as soon as the revision is ready.

Thank you for the response. The budget accounting and the source of the comparison scores are now clarified. The reported filtering improvements also partially address the concern about component contributions.

The main concern remains the evidence for alternating SFT–RL. Comparing two climbing rounds with one does not isolate the benefit of alternation from additional training, and the comparison with continued RL under matched data and compute budgets is still missing. The undisclosed training-data scale, teacher models, and other training details also limit assessment and reproducibility.

Dear Reviewer aAJE:

Thank you very much for your follow-up and for taking the time to consider our responses. The revised manuscript has been submitted. We would sincerely appreciate it if you could reconsider your overall evaluation of the paper. Thank you again for your thoughtful feedback and consideration.

Best regards, The Authors

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

Related papers

Discussion (0)

Sign in to comment.