acceptodds
Under review as a conference paper at ICLR 2027

Beacon: Knowing When and How to Perform Agentic Visual Reasoning

Abstract

The fundamental goal of agentic visual reasoning is to improve the success rate of multimodal large language models (MLLMs) on complex tasks. In this work, we rethink agentic visual reasoning through two key dimensions of tool use: **Mode Adaptiveness** and **Tool Effect**. Mode Adaptiveness characterizes whether an MLLM can recognize when tools are truly necessary and invoke them accordingly, thereby avoiding unnecessary computational overhead while improving performance on challenging problems that require tool assistance. Tool Effect characterizes the actual impact of tool use: tools should extend the model's capabilities on problems unsolvable through tool-free reasoning, while avoiding additional errors on problems that the model can already solve without tools. We conduct a comprehensive analysis to quantify these two properties and empirically reveal that existing agentic visual reasoning models exhibit limited Mode Adaptiveness, while the gains produced by tool use on hard examples are largely offset by the harm introduced on easy examples that the models can already solve. Motivated by these observations, we propose **Beacon**, a novel agentic visual reasoning model trained with supervised fine-tuning (SFT) and reinforcement learning (RL). At the core of Beacon are the *Necessity-Aware Adaptive Reward* and the *Hint-Guided Capability Expansion* mechanism in the RL stage. *Necessity-Aware Adaptive Reward* encourages tool-free solutions when they succeed while preserving full reward for successful tool use when tool-free rollouts fail. Hint-Guided Capability Expansion uses verified, answer-free expert hints to recover learning signals from all-wrong rollout groups, aiming at extending the tool-use capability on the hardest problems. Across 13 benchmarks, Beacon achieves the highest average score among the evaluated open-source models and ranks first on 11 benchmarks. On five diagnostic benchmarks, it improves the average tool-available accuracy over its tool-free accuracy by 1.96 points and achieves the largest tool-gain minus tool-harm score (+3.14 points). These results show Beacon's advanced performance, Mode Adaptiveness, and the net benefit of tool use.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multimodal Large Language Models, Agents

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_F5RzOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper addresses a tangible shortcoming of existing agentic visual reasoning models: they invoke tools indiscriminately, such that the gains tools provide on difficult questions are largely offset by the errors they introduce on easy ones. Therefore, the authors propose Beacon, trained on Qwen3-VL-8B-Instruct via a two-stage SFT-then-RL pipeline. The RL stage contributes two mechanisms: (1) Necessity-Aware Adaptive Reward (NAAR, Equation 1), which gives full reward to correct tool-free answers when tool-free reasoning succeeds within the rollout group, discounted 0.25 reward to correct code-assisted answers in the same case, and full reward to code-assisted solutions when tool-free reasoning fails; (2) Hint-guided Capability Expansion (HCE, Section 3.4.2), which uses Gemini 3.1 Pro to generate answer-free hints for all-incorrect rollout groups, re-samples with hints, then removes the hints during optimization to transfer the discovered behavior to the hint-free policy. Across 13 benchmarks, Beacon achieves the highest average score among open-source 8B models, ranking first on 11 benchmarks.

Critical strengths
  1. By formalizing Mode Adaptiveness and Tool Effect and introducing multi-rollout stability (Section 2.1), the paper provides the community with a principled evaluation protocol applicable to any agentic visual reasoning system. The decomposition of Tool-Gain and Tool-Harm is particularly illuminating, revealing that tool use in prior models is essentially zero-sum (Figure 3c). Even if the specific training method is surpassed, this analytical contribution will remain useful.
  2. The group-conditioned online labeling scheme (Equation 1) avoids the pitfalls of prior adaptive rewards documented in related work. Table 4 shows that NAAR alone improves MA_mean from 56.30 to 59.68 and delta-TE from +1.39 to +2.54.
Critical weaknesses
  1. Gemini 3.1 Pro is used at every stage: generating 15K SFT trajectories, refining trajectories, producing hints for all-incorrect groups during RL, and judging answers at evaluation time. No experiment isolates the contribution of NAAR and HCE from the gains of distilling a frontier model. For instance, would NAAR still help if the SFT data came from Qwen3-VL-72B instead of Gemini? The SFT-only model already achieves 56.91 Overall Acc (Table 4), surpassing all baselines except Metis, suggesting that Gemini-quality SFT data accounts for a substantial proportion of the improvement.
  2. The full NAAR+HCE framework improves Overall Acc from 56.91 (SFT) to 58.98 (+2.07 points, Table 4), a relatively modest margin. Furthermore, Beacon's MA_text is 22.91% (Table 3), meaning it invokes tools on over three-quarters of text-easy questions. The narrative of "adaptive tool invocation" overstates the actual behavior—a more accurate characterization would be "a tool-using model with high tool efficacy."
Questions for authors
  1. How sensitive are the results to the 0.25 discount factor in NAAR (Eq. 1)? Have you tried 0.0, 0.5, or 0.75? A sweep over this parameter is needed to understand whether NAAR is robust or requires careful tuning.
  2. Can you provide results where Gemini 3.1 Pro is replaced by an open-source model (e.g., Qwen3-VL-72B or InternVL-78B)? This would clarify whether the proposed framework's benefits are contingent on access to a frontier proprietary model.
Reviewer_pvn8OpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies agentic visual reasoning for MLLMs and argues that current models use visual tools poorly. It formalises two diagnostic axes, Mode Adaptiveness and Tool Effect, to show that baselines' net tool gain is near zero. Beacon adds two mechanisms: NAAR, which rewards a correct tool-free answer over a redundant code answer, and HCE, which uses strong-teacher, answer-free hints to recover learning signal from all-wrong groups.

Critical strengths
  1. Adaptive tool use is a real, under-explored bottleneck in agentic MLLMs. Framing it as MA and TE is a clean, reusable metric.
  2. NAAR and HCE target two distinct failure modes and are simple, well-motivated additions to GRPO.
  3. Evaluation spans 13 benchmarks plus dedicated MA/TE diagnostics, ablations, training curves and qualitative cases.
Critical weaknesses
  1. This paper reports no tool-call count, token count or latency for Beacon vs baselines vs tool-free. Without a cost/accuracy trade-off, the “knowing when” benefit is only partly demonstrated.
  2. Beacon relies on Python for tool execution, but the generalizability across different types of Python operations remains unclear. For instance, if certain operations are absent from the training data, can the model still generalize to them during testing?
Questions for authors

Please report inference-time cost: mean number of tool calls, tokens and latency, for Beacon vs baselines vs tool-free, on the diagnostic set.

Reviewer_e1hyOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates two limitations of agentic visual reasoning: poor adaptation in deciding when to use tools and limited gains from tool-assisted reasoning. The authors introduce two diagnostic metrics, Mode Adaptiveness and Tool Effect, and propose Beacon, an SFT+RL framework with two components: Necessity-Aware Adaptive Reward (NAAR), which encourages selective tool use, and Hint-Guided Capability Expansion (HCE), which uses expert-generated hints to learn from difficult examples. Beacon achieves the best average performance among evaluated open-source models across 13 benchmarks, with improved tool-use adaptiveness and effectiveness.

Critical strengths
  1. The paper provides a useful framework for evaluating tool use beyond aggregate accuracy.

The distinction between Mode Adaptiveness and Tool Effect addresses an important limitation of existing evaluations: a model can achieve high accuracy without learning when tools are useful. The analysis across multiple agentic visual reasoning models shows that tool availability does not necessarily translate into meaningful performance gains.

  1. The proposed RL framework addresses two concrete shortcomings of conventional tool-use training.

NAAR provides a simple way to balance task correctness with selective tool invocation, rather than uniformly penalizing tool calls. HCE tackles a different problem: difficult samples that produce all-wrong rollout groups provide little useful learning signal under standard GRPO. The proposed approach uses verified expert trajectories to guide exploration and then optimizes the policy without including the hints in its input.

  1. The experiments provide encouraging evidence for the effectiveness of the proposed training components.

Beacon achieves the highest open-source accuracy on 11 of 13 benchmarks, including meaningful improvements on ChartQAPro and GameQA. The ablations in Table 4 also help distinguish the contributions of the training stages: SFT improves average accuracy from 52.91% to 56.91%, while adding standard GRPO yields only 57.10%.

Critical weaknesses
  1. Tool Effect does not establish causal benefits from tool use.

The proposed metrics count correct or incorrect tool-assisted responses without determining whether the tool outputs actually changed the outcome. Model-dependent difficulty partitions also complicate comparisons across methods. Using cost / latency could be potentially better metrics here. This limits how strongly the results support the paper's central claim about genuine capability gains from tools.

  1. Limited evidence of efficient tool selection.

Beacon still invokes tools on approximately 77% of text-easy responses, despite achieving the highest average adaptiveness score. Without latency, token usage, or execution-cost measurements, it is unclear whether the method improves the accuracy-efficiency pareto.

  1. HCE's contribution is difficult to separate from additional teacher supervision.

The method relies on Gemini 3.1 Pro to generate verified hints for difficult problems. Comparisons against simpler teacher-supervised alternatives under matched computational budgets would help establish the specific benefit of hint-guided optimization.

Questions for authors
  1. Can the authors provide a more direct evaluation of the causal contribution of tool execution?

The proposed Tool-Gain and Tool-Harm metrics are useful descriptive statistics, but they do not necessarily measure whether tools change the correctness of an answer. Could the authors evaluate Beacon under controlled settings where tool observations are removed or replaced, while keeping the reasoning prompt and inference budget comparable? This would help establish how often the tool outputs themselves lead to answers that would otherwise be incorrect.

  1. How does Beacon compare with existing methods under a fixed inference budget?

Beacon invokes tools frequently, including on text-easy examples. Could the authors report the generated tokens/cost, and inference latency, and compare accuracy-cost trade-offs against Metis and other baselines? It would be particularly useful to know whether Beacon's advantage persists when models are constrained to comparable inference costs, rather than artificial metrics.

  1. How much of HCE's improvement is attributable to the hint-guided training mechanism rather than additional expert-generated supervision?

Could the authors compare HCE with direct distillation of verified Gemini trajectories or additional rejection-sampled training, using a comparable teacher and training budget? Reporting the metrics would help clarify both the effectiveness and cost of HCE.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.