acceptodds
Withdrawn from ICLR 2027

Seer: Guiding Small Deep-Research Agents with Shared Decision Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Deep-research agents repeatedly decide what to search, read, and verify, often using large cloud-hosted language models for action proposal. However, privacy and local control motivate running agents on users' devices, where limited memory and compute favor smaller models and complicate on-device training. Yet inference-time guidance for these agents faces three obstacles: search through real execution adds tool calls, prompted self-evaluation relies on their judgment, and model-specific guidance risks repeated adaptation costs. We introduce Seer, an inference-time framework providing local guidance that separates action proposal by the agent's model from learned consequence prediction and value estimation. From a shared corpus of real research trajectories, we train two models: a tool-response simulator on observed call–response pairs and a value estimator on histories labeled with discounted terminal correctness. The simulator and estimator are trained offline and reused across agents without further training. Moreover, Seer executes only the selected action and adds only real tool responses to the committed history, avoiding tool calls on discarded candidates at additional inference cost. Across multiple deep-research benchmarks, Seer substantially improves agent performance for different language models. For instance, GAIA accuracy rises from 13.6% to 41.7% for Qwen3.5-4B and from 22.3% to 50.5% for Qwen3.5-9B. Furthermore, our analyses suggest that broader candidate coverage can compensate for imperfect simulation, while prediction and scoring errors limit the benefits of wider imagined search. These findings highlight the potential of reusable models for prediction and evaluation to strengthen small research agents through additional local computation without real execution of discarded candidates.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Deep Research Agent, Edge Computing

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_g99COpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper first explains why deploying deep-research agents on-device is important and analyzes the constraints faced by existing on-device deep-research agents, namely, "local computational resources and external tool-use costs." Building on an analysis of related work, the authors identify the key challenge as "providing reusable guidance that strengthens small agents without model-specific retraining or real execution of discarded candidates." Based on this motivation, the authors propose Seer, an inference-time framework that improves deep-research performance by "decoupling action proposal from consequence prediction and value estimation to enable reusable guidance across agents."

Regarding the methodology, Seer first uses an agent to propose multiple candidate actions. A simulator then simulates the execution of these actions to obtain observations, and a value estimator evaluates how beneficial history is for completing the task. In the experimental section, the authors evaluate Seer on five deep-research agent benchmarks and compare it with three representative baselines, demonstrating effective performance improvements. The authors also conduct several ablation studies and scaling analyses to validate the effectiveness of certain design choices. In terms of efficiency, Seer also demonstrates improvements in tool-use costs.

Critical strengths
  • The paper is clearly structured, forming a coherent logical progression from describing the problem to be addressed and its motivation, through the design of the proposed method, to the experimental validation of its effectiveness.
  • The experiments demonstrated Seer's strong performance across a wide range of models and benchmarks.
Critical weaknesses

1. Lack of a Direct Baseline for Action Value Prediction

The overall design of the proposed method uses a simulator to predict the outcomes of different candidate actions, and then employs a value estimator to evaluate the value of these outcomes toward achieving the task objective. However, I think a straightforward baseline is missing: what if the simulator in the agent system were used to directly predict the final value of each candidate action? How would this affect performance?

2. Lack of Ablation Studies on the Value Estimation Design

Section 4.2 introduces a discount factor for value estimation. Although the ablation studies examine the overall contribution of the value estimator, there is a lack of analysis of specific algorithmic design choices, such as what value of the discount factor would be most appropriate. I have not seen any analysis regarding "Performance Sensitivity to the Discount Factor."

Questions for authors

Figure 3 of the paper compares the tool-call costs of Seer with those of several baselines, and the introduction of the simulator indeed significantly reduces tool-call costs. However, Seer also introduces two additional models, the simulator and the value estimator. Compared with the baselines, how much additional overhead does Seer introduce in terms of metrics such as latency and memory usage? Would this affect the practical usability of Seer?

Reviewer_5aWBOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Seer separates proposing an agent's next action from predicting its consequences and estimating its value. A frozen agent proposes candidate search or webpage-visit actions. An additionally trained tool-response simulator predicts what each candidate might return, and a value estimator scores candidate histories. The system executes only the selected real tool call and appends only the real response to the next interaction history. It spends additional local inference to compare more candidates without actually executing all discarded branches.

Both auxiliary components are initialized from Qwen3.5-9B and trained offline using LoRA before being reused across agents. Keeping the agent model frozen does not mean that the overall system requires no training. Although the main text describes PUCT search, the reported experiments use imagined search of depth one and therefore do not validate benefits from deeper planning.

Critical strengths
  1. Reuse across agents is practically meaningful. The two auxiliary components need not be retrained for each deployed agent, and simulated branches are not mixed into the actual evidence history. The design clearly distinguishes predictions used for action selection from real tool responses that can support citations.

  2. The empirical results are substantial: small models show sizable gains across multiple models and five benchmarks. The paper also includes component additions, candidate-width analyses, and real-tool-call analyses, and discloses the simulator's loss of accuracy. Training questions are separated from evaluation questions, which is a useful strength of the evaluation protocol.

Critical weaknesses
  1. The cost-saving case lacks strong practical alternatives and end-to-end resource comparisons. The central motivation concerns small models on edge devices, but deployment also requires two auxiliary components initialized from a 9B model. Base-model sharing, quantization, KV caches, and batching all affect actual memory use. The main table does not match total generation compute, wall-clock time, or memory, while Figure 3 primarily reports real tool calls for three Qwen agents on BC-200. More importantly, the paper does not compare against a strong public scaffold paired with the same small model as a practical cost-saving alternative. Improvements in planning, retrieval, context management, and verification may also be competitive under the same budget, but that possibility is neither established nor ruled out. A comparison of original ReAct, a strong scaffold, ReAct plus Seer, and the strong scaffold plus Seer, with the backbone and real tool backend fixed, should report quality at matched total budgets or cost at matched target quality. Auxiliary inference, webpage extraction, and amortized offline training should be included. Appendix H acknowledges the absence of hardware measurements for peak memory, energy, and latency. Fewer real tool calls for discarded branches do not by themselves imply lower total resource consumption.

Quantized deployment is particularly under-validated. Appendix H lists quantization of the agent and both auxiliary models as future work, without before/after accuracy, decision-ranking, peak-memory, or end-to-end latency measurements. The complete system, not just the agent, should be compared under a common resource budget against deployable quantized models or a stronger quantized model with a strong scaffold.

The paper also lacks publicly available, similarly sized specialized Deep Research models as practical alternatives. AgentCPM-Explore 4B, MiroThinker-v1.0-8B, and Marco-DeepResearch-8B are small-model candidates, while Tongyi DeepResearch and OpenSeeker cover approximately 30B settings. At least one representative appropriate to the task and experimental cutoff should be considered, comparing a specialized model alone with a general backbone plus Seer, and testing whether Seer adds value to the specialized model. Available specialized weights can be deployed directly; the cost of retraining every model does not exclude that alternative. Historical training costs for both system types should still be included or amortized consistently. Several of these works are already cited, so the gap concerns controlled experimental comparisons, not wholesale omission of the literature. Any tool or prompt adaptation should be documented, with search/web backends, context, real-tool budgets, and total inference costs controlled. Published scores from different papers do not establish a fair winner.

  1. Evaluation primarily covers retrieval-based multi-step question answering and provides insufficient evidence for broader Deep Research capabilities. Final-answer accuracy on GAIA and BrowseComp can validate information seeking, but does not directly evaluate research-insight coverage, cross-document synthesis, citation support, reasoning completeness, or long-report quality. ResearcherBench, DeepResearch Bench, or ResearchRubrics could provide a complementary evaluation; one suitable benchmark would suffice, rather than requiring all of them. Appendix H explains the focus on QA to obtain unambiguous terminal supervision and leaves report-generation tasks to future work. This is a disclosed but unresolved validation boundary. The improvement with S32 also coincides with broader candidate coverage. The distribution shift from training estimators on real histories to scoring simulated histories, and stability across widths, require more controlled evidence. The current results support gains in the tested QA settings, not general effectiveness for open-ended research reports.

Model coverage is also concentrated in two families. Table 2's six configurations include only Qwen and Gemma; the two larger models added in Table 3 are also Qwen, and both the simulator and estimator are initialized from Qwen3.5-9B. Reuse across Qwen sizes or versions is not equivalent to validation across more independent model families. Gemma does provide some cross-family evidence, but transfer to a broader agent distribution remains underexplored. Additional families available before the experimental cutoff and compatible with the deployment budget, such as appropriately sized DeepSeek or GLM models, would strengthen the evidence; alternatively, the generalization claims should be narrowed. The absence of a particular latest model does not invalidate the method, and models released only after submission should not be mandatory comparisons.

Questions for authors
  1. With the same agent backbone and real search/web backend, please compare original ReAct, a strong public scaffold, ReAct plus Seer, and the strong scaffold plus Seer. Match total inference budgets or report quality versus total cost to establish whether Seer adds value to a stronger execution workflow.

  2. Please report residency/base sharing, quantization, peak memory, GPU time, end-to-end latency, real-tool fees, and offline-training amortization for the two 9B auxiliary components. Does visit invoke another language model for evidence extraction? What are the underlying framework, complete prompts, tool implementations, and versions?

  3. Can you evaluate insight coverage, citation/factual support, and synthesis quality on at least one appropriate report-oriented benchmark, such as ResearcherBench, DeepResearch Bench, or ResearchRubrics? If the task scope is not expanded, please explicitly restrict the main conclusions to the tested retrieval-QA capabilities.

  4. Were candidate width, simulation count, and related settings selected on an independent validation set? Please provide repeated-run and paired per-task results for the main metrics, and separate the contributions of the estimator, simulator, and additional candidate coverage under matched compute budgets.

  5. Please specify the experimental and model-selection cutoff. Without retraining the auxiliary models, can you add other independent model families that were publicly available before that cutoff and comparable in hardware requirements, and compare a newer backbone alone, with a strong scaffold, and with Seer? Please also assess performance and whole-system memory/latency before and after quantizing the agent, simulator, and estimator, distinguishing within-family upgrades, cross-family transfer, and quantization robustness.

  6. Can you include at least one publicly available, similarly sized specialized research model, such as AgentCPM-Explore 4B, MiroThinker-v1.0-8B, or Marco-DeepResearch-8B, both alone and with Seer? Tongyi DeepResearch or OpenSeeker could be considered for approximately 30B settings. Use consistent task splits, real tool backends, and budgets, and report adaptation changes, total memory/latency, and tool fees. AgentCPM-Report 8B is another candidate for open-ended reports, but should be evaluated on report-oriented tasks rather than comparing its original report scores with QA accuracy.

Reviewer_u65SOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces Seer, an inference-time guidance framework for deep-research agents that decouples action proposal from tool-response prediction and history-level value estimation. The authors train a shared tool-response simulator and value estimator offline on real research trajectories and reuse them across different frozen agents without agent-specific retraining. At inference, candidate actions are evaluated using predicted tool responses, but only the selected action is executed and only real observations enter the committed history. The experiments use depth-one search and evaluate five question-answering benchmarks across several agent families and scales.

Critical strengths

The separation of action proposal from shared response prediction and value estimation allows the same decision models to guide different agents without retraining each one. Executing only selected actions and retaining only real responses as evidence also provides a meaningful direction for reducing real tool calls during candidate exploration.

Critical weaknesses
  1. Fewer real tool calls do not establish the overall efficiency trade-off. Figure 3 compares accuracy against real tool-call counts on BrowseComp-200, but the simulator, value estimator, and wider search add computation. Total input/output tokens, end-to-end time, and the memory footprint of the auxiliary models are needed to assess the claimed practical benefit under comparable deployment budgets. A feasible context-compression or summarization baseline under the same tools, task settings, and resource limits would also clarify the incremental benefit. This does not require adopting a larger or million-token-context model; context capacity and compression strategy should be distinguished.
  2. The experiments do not isolate the value of generating simulated observations over learned action ranking without observations. Table 4 compares real-response scoring with simulated-response scoring; the latter loses accuracy on six of eight agents at width four. A trained history-and-action scorer selecting among the same proposals without candidate execution or response generation is absent. Such a control would also avoid discarded real tool calls. The full system is useful, but its results do not establish whether the simulator is necessary for the trade-off or whether most benefit comes from the outcome-trained estimator and broader proposal sampling.
Questions for authors
  1. Could the authors report total input/output tokens, end-to-end time, and auxiliary-model memory use for Seer, BoN@4, and AV@4 under comparable tools and deployment budgets? If exact budget matching is infeasible, a quality-cost curve with the unmatched resources stated explicitly would help establish the overall trade-off.
  2. With the same proposals and comparable training data and inference budget, how does a learned history-and-action scorer without observation generation compare with Seer? A limited ablation on an existing benchmark would clarify whether simulated responses contribute beyond learned ranking and proposal sampling.
  3. Is there an existing or feasible context-compression/summarization baseline within the same tools and resource limits? If so, does Seer improve the accuracy-cost trade-off over it?
  4. What explains the model-dependent width gains in Table 2? On GAIA, S4 to S32 improves Qwen3.5-9B from 35.9% to 50.5% and Gemma-4-e4b-it from 24.3% to 36.9%, but Gemma-4-e2b-it only from 30.1% to 31.1%, with regressions in some other settings. Could existing prediction-error or ranking diagnostics explain this variation?
Reviewer_YiSfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Seer, an inference-time guidance framework for improving frozen deep-research agents. The method decouples action proposal, tool-response prediction, and history-level value estimation: the agent generates candidate actions, a simulator predicts their tool responses, and a value estimator scores the resulting imagined histories. Only the selected action is executed, and only real tool responses are added to the committed interaction history. The two models are trained on real research trajectories collected from multiple agents and reused across different agent models. Experiments cover several information-seeking benchmarks, model families, and model scales, showing substantial accuracy improvements and lower real tool-call costs than methods that execute multiple complete trajectories.

Critical strengths
  1. The problem is practically relevant, and the system design is clear. Improving research agents’ action selection without updating their underlying models is a worthwhile direction. The paper explicitly distinguishes predicted content from real evidence and prevents simulated tool responses from entering the committed history. This design makes model errors primarily affect action selection rather than directly introducing simulated retrieval evidence into the real interaction history.

  2. The ablations reveal a useful performance–cost trade-off. Table 4 shows that the learned estimator generally outperforms prompted scoring, while replacing real observations with predicted responses can reduce accuracy. Increasing candidate coverage sometimes compensates for this loss. The authors also report that larger candidate widths do not consistently improve performance, rather than presenting only monotonically improving results.

Critical weaknesses
  1. The paper does not sufficiently establish the value of explicit tool-response generation over simpler action-scoring alternatives The central design is to predict tool responses and then evaluate the resulting histories. However, the experiments do not adequately distinguish this approach from a direct alternative with action-scoring.

  2. Applicability to stronger frontier-level agents remains unclear Although Table 3 reports improvements on larger Qwen models, the evaluation remains limited to the Qwen and Gemma families. It remains unclear whether comparable gains would extend to other frontier-level agents with stronger native planning and tool-use capabilities, or whether the shared 9B simulator and estimator would become a bottleneck in such settings. A representative evaluation on a stronger agent outside the tested families, reporting both accuracy improvements and additional inference costs, would help establish the scope and practical value of the approach.

Questions for authors

1.Is explicit tool-response simulation necessary? 2.Do the benefits extend beyond the evaluated model families?

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

Related papers

Discussion (0)

Sign in to comment.