acceptodds
Under review as a conference paper at ICLR 2027

Switch-Reasoner: Learn When to Think in Multitask Mixtures via Reinforcement Learning

Abstract

Multimodal Large Language Models (MLLMs) often follow a fixed Think-then-Answer paradigm, which is inefficient in heterogeneous multitask settings because simple inputs may not require explicit reasoning while difficult ones can benefit substantially from it. Learning when to think is also unstable during post-training, where imbalanced rollouts can drive the model toward always-thinking or always-direct behavior. We propose Switch-Reasoner, a GRPO-based framework that learns to adaptively select reasoning modes for MLLMs. It treats thinking as a virtual tool invocation and allows the model to either answer directly or invoke explicit reasoning before answering. To stabilize this decision, we introduce a dual-level regulation mechanism that balances the overall use of Thinking Mode and Direct Mode while providing sample-level supervision based on the relative benefit of the two choices. Experiments on 11 multimodal tasks show that Switch-Reasoner reduces unnecessary reasoning while maintaining strong performance, achieving a better accuracy-efficiency trade-off.

Primary area:
foundation or frontier models, including LLMs
Keywords:
MLLMs, reasoning, efficiency

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_PfBQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Switch-Reasoner, which treats thinking as a tool and adaptively selects reasoning modes for MLLMs. Specifically, Switch-Reasoner has global mode balance control and sample-level fine-grained optimization. This dual-level regulation mechanism balances the Thinking and Direct Modes.

Critical strengths
  1. This paper proposed a dual-level regulation mechanism to balance the overall use of Thinking Mode and Direct Mode. It's somewhat new.
Critical weaknesses
  1. This paper fails to provide a sufficiently detailed analysis. Somehow, Think Mode is only a format, and even when the model selects the Direct Mode, it may still present a lengthy reasoning process in Chain-of-Thought (CoT) form. In addition to reporting the "think rate", I suggest that the authors provide statistics on output length, as this is crucial.
  2. To the best of my knowledge, the "thinking-as-tool" concept was first mentioned in the Anthropic blog: https://www.anthropic.com/engineering/claude-think-tool, which has been widely discussed within the community. There are also several highly relevant works that the authors should discuss in the "Related Work" section: https://arxiv.org/html/2505.13379v2; https://arxiv.org/html/2508.12265v2.
  3. The baseline comparisons in this work are insufficient; it only compares against GRPO. It should incorporate response efficiency (response length) and include comparisons with existing work on efficient reasoning.
  4. The ablation study variant removed the sample-level fine-grained correction. What would be the effect if the sample-level fine-grained correction were retained while the global balance control was removed? Additionally, I am particularly concerned about the fact that global balance is heavily influenced by batch sampling; how do extreme sampling outcomes—such as a batch consisting entirely of tasks requiring high inference effort followed by a batch of simple tasks—affect training stability?
  5. Following up on the previous point, I believe that training stability is heavily influenced by the data distribution; therefore, I think this aspect should be described in greater detail in the main text of the paper. What is the composition of your training data, and does it affect the final model's performance and think ratio?
Questions for authors

See Weaknesses.

5 comments

Response to Reviewer PfBQ (1/4)

We sincerely thank the reviewer for the detailed feedback regarding routing behavior, related work, efficiency measurements, and training stability.

W1: Can the Direct mode still generate lengthy reasoning processes?

Thank you for highlighting this important distinction.

The operational definition of Direct is generating an answer without invoking the virtual thinking tool. It does not semantically guarantee that the model performs no internal reasoning or generates no explanatory text.

To examine whether Direct actually produces shorter outputs, we analyzed the generated token distributions of a supplementary 4B checkpoint trained for 110 steps.

The analysis covers 5,500 evaluation samples. Token counts include both model-generation stages and model-generated tool calls. Injected tool-response context is recorded separately and excluded from the generated-output count.

ModeSamplesMean TokensMedian Tokens95th PercentileMaximum Tokens
Direct1,43516.21132644
Thinking4,065305.251361,4362,065

The Direct branch generates only 16.21 tokens on average, with a maximum of 44 tokens among the evaluated samples.

In contrast, the Thinking branch generates 305.25 tokens on average and exhibits a substantially longer tail.

These measurements demonstrate that Direct outputs are indeed concise in the evaluated implementation.

We will clarify that the thinking rate measures the frequency of explicit thinking-tool invocation, whereas generated token statistics provide an independent and more direct characterization of actual output length.

W2: Thinking-as-Tool has precedents, and the related work should be expanded.

We agree with the reviewer and appreciate this suggestion.

We will expand the related work section to discuss representative studies on tool-based reasoning and adaptive reasoning allocation, including:

  1. Anthropic's think tool: Provides a precedent for organizing additional reasoning through an explicit tool interface.
  2. Thinkless: Learns to select between shorter and longer reasoning behaviors using dedicated control tokens and mode-aware reinforcement learning.
  3. Fast, Slow, and Tool-augmented Thinking for LLMs: A Review: Provides a broader overview of adaptive reasoning and tool-augmented reasoning paradigms.

These studies demonstrate that explicit reasoning control and reasoning-tool interfaces are not new concepts.

Our work builds upon these foundations while focusing on a different training challenge: stabilizing adaptive reasoning-mode selection during GRPO optimization across heterogeneous multimodal tasks.

Specifically, our contribution integrates:

  • Explicit Thinking/Direct routing actions.
  • A delayed Global controller for batch-level mode regulation.
  • Online sample-level counterfactual routing supervision.

We will revise our contribution statement to emphasize the dual-level training mechanism rather than claiming novelty in the tool abstraction itself.

Response to Reviewer PfBQ (2/4)

W3: Additional efficiency statistics and comparisons with existing efficient reasoning methods are needed.

Thank you for this suggestion. We have incorporated thinking rate, generated token count, and end-to-end latency into a unified comparison, and added two adaptive reasoning baselines, Thinkless/DeGRPO and AdaptThink.

MethodAccuracyThinking RateAvg. Generated TokensAvg. Latency (s)
Switch-Reasoner (Full)71.6051.53%142.552.94
Thinkless/DeGRPO59.020.00%11.760.35
AdaptThink62.2312.33%65.260.73

The full Switch-Reasoner achieves an accuracy of 71.60, outperforming Thinkless/DeGRPO and AdaptThink by 12.58 and 9.37 percentage points, respectively.

Meanwhile, Thinkless/DeGRPO exhibits a thinking rate of 0.00%, indicating complete collapse into Direct. Although AdaptThink does not completely collapse, its thinking rate of only 12.33% indicates a pronounced preference for Direct.

In contrast, our method keeps both Direct and Thinking active, avoiding the substantial accuracy degradation associated with reducing computational costs primarily by suppressing reasoning.

We also acknowledge that the two baselines achieve lower token consumption and latency. Therefore, Switch-Reasoner does not dominate all efficiency metrics.

A within-model comparison between default routing and forced Thinking provides a more direct measurement of the efficiency gains from routing.

Default routing achieves an accuracy of 71.60 with 142.55 generated tokens and 2.94 seconds of latency, whereas forced Thinking requires 251.45 tokens and 5.49 seconds to achieve an accuracy of 71.80.

This corresponds to a 43.3% reduction in generated tokens and a 46.4% reduction in latency, with only a 0.20-percentage-point accuracy gap.

This experiment provides more direct evidence that the dual-level regulation mechanism reduces unnecessary reasoning while maintaining high accuracy.

Response to Reviewer PfBQ (3/4)

W4: What happens when Global regulation is removed? Does extreme alternation between task groups affect training stability?

Thank you for raising this question.

We address it through two complementary analyses: an ablation of Global regulation and a controlled experiment on task-order sensitivity.

Global regulation ablation.

We introduce a Sample-only variant that retains sample-level counterfactual routing supervision while disabling Global regulation and the extreme-imbalance safeguard.

Following the protocol of averaging the last three validation checkpoints, the additional ablation results are as follows:

Method (4B)Overall ScoreΔ vs. Full Model
Switch-Reasoner (Full)71.60—
Sample-only69.82−1.78

Without Global, Sample-only reaches a thinking rate of 99.18%, compared with 51.53% for the full model, indicating a strong preference for Thinking and near-complete mode collapse.

Meanwhile, the overall score decreases from 71.60 to 69.82.

These results support the role of Global in mitigating mode collapse under the evaluated setting, in terms of both routing behavior and task performance.

Task-order sensitivity.

We further investigated whether changes in batch composition affect controller stability.

We fixed the initialization, 3,520 training samples, batch size, random seed, and execution configuration, varying only the presentation order of task groups.

Under the randomly mixed setting, samples from different tasks were interleaved. Under the alternating setting, task groups representing relatively higher and lower reasoning demands were alternated during the first 80 steps, followed by 30 steps of training on lower-demand task groups.

These task groups serve only as proxies for reasoning demand and are not sample-level difficulty annotations.

Both configurations were evaluated on the same 1,100 validation samples.

We define an extreme-mode batch as a training batch with a thinking rate of at least 98% or at most 2%.

Sampling Order (Seed 2)AccuracyThinking RateExtreme-mode Batches
Randomly Mixed Tasks71.6051.53%0/110
Alternating Task Groups69.9665.55%16/110

Under the alternating task-group schedule, both reasoning modes remain active. However, accuracy decreases from 71.60 to 69.96 (a reduction of 1.64 percentage points), and the number of extreme-mode batches increases.

This indicates that task order can affect routing stability even when the overall training-data composition remains unchanged.

We will present the corresponding training dynamics and controller-signal changes in Supplementary Figure S2.

Additionally, we observed that another task-alternation experiment resulted in a final thinking rate of 0%. However, because its execution configuration was not consistent with the randomly mixed control setting, we do not include it in the paired quantitative comparison.

We will clarify the task-mixing protocol and acknowledge that, although Global regulation mitigates mode collapse under the evaluated settings, it does not guarantee stability under arbitrary task sequences.

Reviewer_uX5ZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work addresses the inefficiency of the fixed think-then-answer paradigm under heterogeneous multimodal and multi-task distributions. Simple inputs can often be answered reliably in a single pass, while more challenging questions are the ones that truly benefit from explicit deliberation. The authors observe that, when GRPO post-training is applied to a mixture of tasks, immediate reward advantages and the imbalanced distribution of sampled trajectories can push the policy toward degenerate behavior, causing it to collapse into either always reasoning or always answering directly. To address this issue, they propose Switch-Reasoner, a GRPO-based framework that treats reasoning as an invocable virtual tool and turns the decision of whether to reason into an explicit, learnable action within the reinforcement learning objective.

Critical strengths
  • Well-motivated problem formulation. The work targets a practically relevant limitation of current reasoning pipelines: explicit chain-of-thought incurs unnecessary latency and efficiency loss on easy instances, yet remains beneficial on hard ones. Adapting the reasoning mode to each input offers a natural trade-off between accuracy and computational cost. The problem is stated clearly and is well bounded.
  • A concise and readily deployable "Thinking-as-Tool" formulation. Reasoning is represented as a discrete, observable action mediated through a tool-calling interface, which allows the decision to reason to be optimized directly under reinforcement learning rather than left to prompt-level heuristics. The proposed policy decomposition is principled, carries a modest engineering overhead, and should be straightforward to reproduce.
  • A complementary and internally consistent two-level control scheme. A global controller supplies a batch-level, coarse-grained anti-collapse mechanism that guarantees exploration of both modes, while per-sample counterfactual supervision determines instance-by-instance whether reasoning or a direct answer is appropriate.
Critical weaknesses

The core idea of "learning when to think" offers limited novelty. The general concept of adaptive reasoning has already been extensively studied in large language models. Extending this idea to multimodal GRPO training is more of an incremental application than a conceptual breakthrough. The "thinking-as-a-tool" reformulation is a clever implementation trick, but it does not constitute a methodological innovation.

The paper lacks a direct comparison with existing adaptive-reasoning baselines. It only compares against variants such as Vanilla, GRPO-Thinking (100% thinking), and GRPO-Direct (0% thinking), rather than against prior adaptive-reasoning methods. Without these baselines, it remains unclear whether the proposed dual-layer modulation genuinely outperforms simpler or existing adaptive strategies, or whether the performance gains merely stem from some form of mode switching.

Questions for authors

Comparison with existing methods. Under the same multimodal GRPO setting, how does Switch-Reasoner perform relative to prior "learning-to-think" methods? Does the dual-layer modulation significantly outperform simpler adaptive strategies?

Training cost. Compared with standard GRPO-Thinking, what is the actual increase in training computation, GPU-hours, or wall-clock time?

Quantification of inference efficiency. By how much are the average latency or FLOPs actually reduced?

4 comments

Response to Reviewer uX5Z (1/2)

We sincerely thank the reviewer for the critical and constructive feedback regarding novelty, baseline comparisons, and efficiency evaluation.

W1: The novelty of learning when to think and Thinking-as-Tool is limited.

Thank you for raising this concern. We agree that adaptive reasoning and tool-based organization of reasoning have been extensively explored in prior work.

Our technical contribution is neither the Thinking/Direct decision itself nor the concept of exposing reasoning to the model through a tool interface.

Instead, Switch-Reasoner addresses a specific optimization challenge: learning a stable, input-adaptive reasoning policy during GRPO training on heterogeneous multimodal task mixtures.

We identify two interacting sources of difficulty.

First, at the batch level, transient reward asymmetries and imbalanced mode-selection frequencies can drive the policy toward excessive use of either Thinking or Direct.

Second, at the sample level, a single global preference cannot adequately capture the varying reasoning requirements of individual multimodal inputs.

To address these challenges, we introduce a dual-level regulation mechanism:

  • Global Mode Balance Control: A delayed-feedback controller adjusts reward contributions based on mode usage and relative correctness rates, together with a safeguard against extreme imbalance.
  • Sample-level Fine-grained Optimization: Online counterfactual comparisons identify inputs for which reasoning provides clear benefits and inputs for which Direct is already sufficiently reliable.

Thinking-as-Tool provides an explicit routing interface through which these complementary training signals can influence mode selection.

We will further refine our novelty claims to emphasize this combined mechanism, acknowledge prior contributions to adaptive reasoning and thinking-tool approaches, and avoid presenting the routing interface itself as the primary conceptual innovation.

W2 / Q1: Does the proposed method outperform simpler switching strategies and existing adaptive methods?

MethodAccuracyThinking RateAvg. Generated TokensAvg. Latency (s)
Switch-Reasoner (Full)71.6051.53%142.552.94
Thinkless/DeGRPO59.020.00%11.760.35
AdaptThink62.2312.33%65.260.73

This comparison reveals mode collapse or a pronounced mode-selection bias in existing adaptive methods under our training setting.

Thinkless/DeGRPO exhibits a thinking rate of 0.00%, indicating complete collapse into Direct. AdaptThink achieves a thinking rate of only 12.33%; although it does not completely collapse, it exhibits a strong preference for Direct and insufficiently activates the Thinking branch.

Correspondingly, their accuracies are 59.02 and 62.23, respectively, which are 12.58 and 9.37 percentage points lower than Switch-Reasoner's 71.60.

In contrast, Switch-Reasoner maintains effective use of both modes with a thinking rate of 51.53%, achieving substantially higher accuracy. This is consistent with the objective of our dual-level regulation mechanism: mitigating domination by a single mode while optimizing routing at the sample level.

We acknowledge that the two baselines achieve lower token consumption and latency because they select Direct more frequently. Thus, the comparison reflects different accuracy-efficiency trade-offs. Moreover, additional sampling overhead may differ across methods, and the same number of RL steps does not necessarily imply equal training compute.

Response to Reviewer uX5Z (2/2)

Q2: What is the additional training cost compared with GRPO-Thinking?

Our method introduces additional computational overhead through counterfactual sampling.

For each input, both methods generate G=8G=8 policy rollouts. In addition, Switch-Reasoner generates K=4K=4 forced-Direct rollouts and K=4K=4 forced-Thinking rollouts to estimate the relative benefits of different routing decisions.

Therefore, Switch-Reasoner processes additional trajectories during training, but these counterfactual samples are not required during inference.

To quantify the overhead, we conducted a controlled 20-step timing comparison. Both configurations started from the same initial model and used the same input order, a batch size of 8, and two NVIDIA H20 GPUs.

We report the average step time over iterations 6–19 to characterize steady-state training performance. Total wall-clock time includes initialization, training, final validation, and checkpoint saving.

ConfigurationSteady-state Time (s/step)Relative Step TimeTotal Time (min)GPU-hours
Switch-Reasoner (Full)101.731.45×42.031.422
GRPO-Thinking70.091.00×29.420.981

We acknowledge that this represents a non-negligible training cost. The corresponding trade-off is that additional training-time supervision enables the model to learn an adaptive reasoning policy, thereby avoiding unnecessary reasoning after training.

The GPU-hours reported here correspond only to this 20-step measurement and should not be interpreted as the total cost of the full training experiments in the paper.

Q3: How much inference latency or FLOPs are actually reduced?

We agree that computational efficiency should be evaluated using actual generation costs rather than inferred solely from the frequency of Thinking-mode usage.

The original paper reports that the fully trained 4B Switch-Reasoner achieves a thinking rate of 51.53%, compared with 100% for GRPO-Thinking, while improving the overall score from 70.66 to 71.60.

To quantify the actual inference savings from adaptive routing, we compare default routing with forced Thinking (bias +24) using the same Switch-Reasoner model on the same 1,100 evaluation samples.

The average generated token count includes both model-generation stages and tool-call tokens, while latency measures end-to-end inference cost.

ConfigurationAccuracyThinking RateAvg. Generated TokensAvg. Latency (s)
Default Routing71.6051.53%142.552.94
Forced Thinking71.80100.00%251.455.49

Compared with forced Thinking, default routing reduces average token consumption by 43.3% and average end-to-end latency by 46.4%, with only a 0.20-percentage-point decrease in accuracy.

Even when all inputs are forced into Thinking, the accuracy improvement is not commensurate with the increase in computational cost.

These results demonstrate that the efficiency benefits of Switch-Reasoner extend beyond a lower tool-call frequency and are directly reflected in shorter generations and lower actual request latency.

We emphasize that this comparison is conducted by modifying the routing bias of the same model, rather than directly comparing the latency of default routing against an independently trained GRPO-Thinking model.

We also do not equate the thinking rate directly with FLOPs. Instead, we report actual generated token counts and end-to-end latency as efficiency metrics.

Thank you for the detailed responses and additional measurements, which have partially clarified the efficiency and training-cost questions. However, my central concern about methodological novelty remains insufficiently addressed: the substantive technical advance over prior adaptive reasoning and routing approaches—balance reward shaping, counterfactual sample comparison, and tool-based mode routing—is still unclear. The additional comparisons show some accuracy and efficiency gains, but the baseline evaluation may reflect unequal tuning conditions, and the efficiency comparison is made against a forced version of the same model rather than an independently trained baseline. These results are therefore insufficient to change my assessment. I maintain my rating.

Thank you for your further feedback. We would like to clarify our contribution and the baseline comparison settings.

We designed a dual-level regulation mechanism for complex multitask settings. It combines sample-level tool-benefit assessment with global usage regulation: the former uses counterfactual outcomes from the current policy to determine whether an additional invocation is beneficial, while the latter adjusts training signals based on usage frequency and relative task performance to mitigate excessive preference for a single route. This design extends beyond reasoning-mode selection and can be applied to similar tool-use efficiency optimization scenarios. Providing this reusable optimization framework is the contribution we wish to emphasize.

Regarding baseline fairness, all other adaptive baselines were evaluated under exactly the same experimental settings. Under these shared conditions, our method achieves substantially higher accuracy while avoiding always-on thinking, demonstrating more effective adaptive reasoning.

Thank you again for your comments. We will incorporate these clarifications into the revised manuscript to explain our contribution and experimental settings more clearly. We hope these clarifications will help you reconsider your assessment of our work.

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

Switch-Reasoner treats the decision to reason as an explicit tool-call action within a multimodal, multitask policy. GRPO training combines a global mode-balancing objective with a sample-level reward based on counterfactual routes. On Qwen3-VL-4B and 8B, the method reports aggregate performance comparable to always-thinking models while invoking reasoning on substantially fewer examples.

Critical strengths

Switch-Reasoner treats the decision to reason as an explicit tool-call action within a multimodal, multitask policy. GRPO training combines a global mode-balancing objective with a sample-level reward based on counterfactual routes. On Qwen3-VL-4B and 8B, the method reports aggregate performance comparable to always-thinking models while invoking reasoning on substantially fewer examples.

Critical weaknesses
  1. The reported accuracy trade-off is not sufficiently transparent. At 4B, mathematics falls from 71.90 to 67.23 and grounding from 61.40 to 60.27 relative to GRPO-Thinking, while the overall score increases by 0.94 points. These results can coexist under a valid aggregation rule, but the rule and all eleven task scores are needed to interpret the gain. Without them, it is difficult to assess whether the method preserves performance broadly or trades losses in some tasks for gains in others.
  2. Reduced thinking frequency does not establish reduced inference cost. Selected examples may generate longer reasoning traces, and routing introduces its own overhead. Mean generated-token counts and measured latency are therefore necessary to support the efficiency claim. Ideally, the comparison would show an accuracy–cost curve across routing budgets, since a single operating point does not establish that the method allocates computation better than alternatives.
  3. The baselines do not adequately test the adaptive-routing contribution. Always-thinking and never-thinking are useful endpoints, but neither competes on the central decision of when to reason. Comparisons with relevant adaptive methods, or a controlled implementation using the same switchable interface, would clarify the benefit of the proposed rewards. The relationship to AdaTooler-V also needs more precise discussion before making a priority claim for multimodal adaptation.
Questions for authors

Please provide the aggregation rule and complete per-task results, together with token counts and measured latency. How does the approach compare with an adaptive baseline at a matched inference budget, and how does performance change as the target thinking rate varies?

3 comments

Response to Reviewer wXJB (1/2)

We sincerely thank the reviewer for the insightful comments regarding cross-task performance, inference efficiency, and the relationship between our method and existing adaptive reasoning approaches.

W1 / Q: Why does the overall score improve despite performance degradation on mathematics and grounding tasks?

Our overall score is the equally weighted average across 11 multimodal benchmarks, each evaluated using 500 samples. Therefore, it reflects aggregate performance across heterogeneous tasks rather than requiring improvements on every individual task.

As shown in Table 3, the category-level results for the 4B model, compared with GRPO-Thinking, are as follows:

CategoryGRPO-ThinkingSwitch-ReasonerΔ
Mathematics71.9067.23−4.67
Chart/Doc91.5291.66+0.14
Grounding61.4060.27−1.13
General64.8468.26+3.42
Overall70.6671.60+0.94

Although performance decreases on mathematics and grounding tasks, Switch-Reasoner improves the General category score by 3.42 points. Importantly, this category covers 6 of the 11 benchmarks.

Consequently, the improvements across the more extensively represented General benchmarks outweigh the performance degradation on mathematics and grounding tasks, resulting in an overall improvement of 0.94 points.

This reflects a cross-task performance trade-off rather than consistent improvements across all tasks. Our objective is to learn a unified adaptive reasoning policy across heterogeneous multimodal tasks, where the computational benefit of explicit reasoning can vary substantially across inputs.

We will explicitly discuss this trade-off in the revised manuscript to avoid implying that Switch-Reasoner improves performance on every benchmark.

W2 / Q: Does the reduced thinking frequency translate into actual computational savings? How does performance vary under different reasoning budgets?

We agree that the thinking rate alone is insufficient to establish inference efficiency. We therefore conducted additional analyses of generated token counts and end-to-end inference latency.

In the original paper, Switch-Reasoner achieves an overall score of 71.60 with a thinking rate of 51.53%, compared with an overall score of 70.66 and a thinking rate of 100% for the independently trained GRPO-Thinking baseline.

We adjust the logit bias associated with the thinking-tool call in the initial generation decision, allowing us to vary the frequency of Thinking without retraining the model. The results are presented below.

Routing ConfigurationAccuracyThinking RateAvg. Generated TokensAvg. Latency (s)
Bias −2466.330.00%12.530.39
Bias −1269.3241.09%123.292.16
Default Routing (Bias 0)71.6051.53%142.552.94
Bias +1271.8197.64%243.435.26
Bias +2471.80100.00%251.455.49

The results show that increasing the frequency of Thinking does not yield proportional accuracy improvements. When the bias increases from 0 to +12, the thinking rate rises from 51.53% to 97.64%, the average generated tokens increase from 142.55 to 243.43, and latency increases from 2.94 to 5.26 seconds. However, accuracy improves only marginally, from 71.60 to 71.81 (+0.21 percentage points). Further increasing the thinking rate to 100% yields an accuracy of 71.80, with no additional improvement.

These findings suggest that increasing explicit reasoning for all inputs does not necessarily provide commensurate benefits.

Compared with forced Thinking (bias +24), default routing reduces the thinking rate from 100% to 51.53%, average generated tokens from 251.45 to 142.55 (a 43.3% reduction), and average end-to-end latency from 5.49 to 2.94 seconds (a 46.4% reduction), with only a 0.20-percentage-point decrease in accuracy.

Conversely, reducing the bias to −12 or −24 further decreases computational costs but also lowers accuracy to 69.32 and 66.33, respectively.

Overall, default routing achieves a favorable accuracy-efficiency trade-off rather than simply minimizing the thinking rate. This comparison isolates inference-time routing within the same model; forced Thinking is not equivalent to the independently trained GRPO-Thinking baseline.

Response to Reviewer wXJB (2/2)

W3 / Q: How can the contribution of adaptive regulation be validated? How does Switch-Reasoner differ from AdaTooler-V?

We agree that the contribution of adaptive regulation should be distinguished from the benefit of merely providing a switchable reasoning interface.

We also agree that existing adaptive methods are important external baselines.

MethodAccuracyThinking RateAvg. Generated TokensAvg. Latency (s)
Switch-Reasoner (Full)71.6051.53%142.552.94
Thinkless/DeGRPO59.020.00%11.760.35
AdaptThink62.2312.33%65.260.73

This comparison reveals mode collapse or a pronounced mode-selection bias in existing adaptive methods under our training setting.

Thinkless/DeGRPO exhibits a thinking rate of 0.00%, indicating complete collapse into Direct. AdaptThink achieves a thinking rate of only 12.33%; although it does not completely collapse, it exhibits a strong preference for Direct and insufficiently activates the Thinking branch.

Correspondingly, their accuracies are 59.02 and 62.23, respectively, which are 12.58 and 9.37 percentage points lower than Switch-Reasoner's 71.60.

In contrast, Switch-Reasoner maintains effective use of both modes with a thinking rate of 51.53%, achieving substantially higher accuracy. This is consistent with the objective of our dual-level regulation mechanism: mitigating domination by a single mode while optimizing routing at the sample level.

We acknowledge that the two baselines achieve lower token consumption and latency because they select Direct more frequently. Thus, the comparison reflects different accuracy-efficiency trade-offs. Moreover, additional sampling overhead may differ across methods, and the same number of RL steps does not necessarily imply equal training compute.

Regarding AdaTooler-V, it regulates rewards for visual-tool usage based on pre-annotated Tool Benefit Scores. In contrast, our method selects between Direct and Thinking through a virtual thinking tool, estimates routing benefits using on-policy counterfactual sampling during training, and combines these estimates with batch-level Global regulation.

We will explicitly acknowledge the shared principle of benefit-aware supervision and position our contribution around the proposed dual-level training mechanism, avoiding overly broad claims of priority in multimodal adaptive reasoning.

We sincerely thank the reviewer for recognizing Switch-Reasoner’s ability to maintain overall performance while reducing thinking invocations. We hope that our clarification of cross-task performance trade-offs, together with additional token and latency measurements, analyses across routing budgets, and adaptive baseline comparisons, has adequately addressed your main concerns. We would greatly appreciate it if you could consider our responses and share any further feedback. Thank you again for your constructive comments and suggestions, which have helped us clarify our contribution and strengthen our experimental evaluation.

Reviewer_G8JsOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies adaptive reasoning in multimodal large language models across heterogeneous tasks. The authors propose Switch-Reasoner, which allows a model to either answer directly or invoke a virtual thinking tool before producing its answer. During GRPO training, a global controller regulates the usage of the two modes, while sample-level comparisons between direct and thinking rollouts provide additional supervision for mode selection. Experiments on 11 multimodal tasks using Qwen3-VL-4B and Qwen3-VL-8B evaluate the method against fixed-mode GRPO baselines and examine its overall performance and thinking rate.

Critical strengths

(i) The paper addresses the important problem of adapting reasoning effort to different multimodal inputs. Avoiding unnecessary explicit reasoning while retaining it for difficult problems is relevant to improving the efficiency of general-purpose multimodal models.

(ii) The proposed framework provides an explicit mechanism for reasoning-mode selection and combines global mode regulation with sample-level supervision.

Critical weaknesses

(i) The sample-level routing rule in Eq. (11) does not account for cases where both modes can answer correctly. A sample is classified as safe-direct only when Direct Mode achieves sufficiently high accuracy and outperforms Thinking Mode by a specified margin. When both modes perform equally well, the sample is classified as uncertain and receives no preference for Direct Mode, even though explicit reasoning provides no additional benefit. This is inconsistent with the paper's motivation of avoiding unnecessary reasoning on simple inputs. The paper does not explain why the routing rule excludes these cases or how they are handled during training.

(ii) The paper does not clearly establish its contribution beyond existing adaptive reasoning methods. Section 2.2 discusses several approaches that already learn when to think, but does not explain why they are insufficient for stable mode selection in heterogeneous multitask GRPO training. The paper identifies mode collapse as a central problem, yet the experiments do not isolate the effect of the proposed Global Mode Balance Control. The fixed Thinking and Direct baselines cannot demonstrate whether a learned switching policy suffers from mode collapse, and the Switch-Reasoner* ablation still retains the global controller. Without a comparison that removes this controller, its contribution to preventing mode collapse remains unclear.

Questions for authors

(i) Please explain why the sample-level routing rule does not favor Direct Mode when both modes achieve similar accuracy, despite the goal of avoiding unnecessary reasoning.

(ii) Please clarify how Switch-Reasoner addresses the limitations of existing adaptive reasoning methods and how the Global Mode Balance Control contributes to preventing mode collapse.

3 comments

Response to Reviewer G8Js

We sincerely thank the reviewer for the constructive feedback. We have revisited our routing supervision design and conducted additional analyses to clarify the individual contributions of the two-level regulation mechanism.

W1 / Q1: When both modes produce correct answers, why not explicitly prefer the Direct mode?

Thank you for raising this important question.

Equation (11) is intentionally designed to provide additional routing supervision only when there is sufficiently clear evidence of a performance difference between the two modes. Specifically, we generate K=4K=4 counterfactual rollouts for each mode to estimate the empirical success rates of Direct and Thinking.

Equal observed success rates do not necessarily imply equal underlying reliability, particularly with a limited sampling budget. Directly labeling such examples as safe-direct could therefore introduce noisy routing supervision and prematurely suppress reasoning.

We instead assign these cases to uncertain, for which the additional sample-level routing reward is zero. This does not exclude these examples from training: correctness rewards, format rewards, and global mode regulation remain active, allowing the model to learn to use Direct when appropriate.

The safe-direct criterion is also deliberately conservative: Direct must achieve a sufficiently high empirical success rate, while Thinking must provide no substantial additional benefit. This asymmetric design avoids prematurely routing reasoning-intensive inputs to Direct.

The ablation in Appendix E, Table 4 provides empirical support for this design. The Safe-direct loose variant allows tied cases meeting the success-rate threshold to be classified as safe-direct. On Qwen3-VL-4B, this reduces the think rate from 51.53% to 13.28%, but also decreases the overall score from 71.60 to 68.32 (−3.28 points). This demonstrates that relaxing the routing criterion can substantially reduce thinking frequency at the cost of overall performance.

We acknowledge that our design does not explicitly favor Direct whenever the empirical success rates are tied. We prioritize the reliability of routing supervision over an unconditional preference for efficiency. We will clarify this trade-off and the motivation for the uncertain category in the revision.

W2 / Q2: What are the contributions of the two-level regulation mechanism? Does Global mitigate mode collapse?

Thank you for this question, which directly concerns the central contribution of Switch-Reasoner.

Our method addresses two complementary challenges in heterogeneous multitask GRPO training.

Global Mode Balance Control adjusts batch-level mode usage based on relative accuracy and routing frequency, preventing transient reward imbalances from disproportionately reinforcing a dominant mode.

Sample-level Fine-grained Optimization provides input-dependent routing supervision by comparing counterfactual Direct and Thinking outcomes generated by the current policy.

These components operate at different levels: Global stabilizes training dynamics, while Sample promotes input-specific allocation of reasoning.

To isolate the contribution of Global, we introduce a Sample-only variant that retains sample-level routing supervision but removes Global regulation and its safeguards against extreme imbalance.

Consistent with Table 1 in the main paper, the following results are averaged over the final three validation checkpoints.

Method (4B)Overall scoreDifference from full modelThink rate
Switch-Reasoner (full)71.60—51.53%
Switch-Reasoner* (w/o Sample)70.79−0.8185.73%
Sample-only (w/o Global)69.82−1.7899.18%

Removing Sample increases the think rate from 51.53% to 85.73% and reduces the overall score by 0.81 points. With Sample retained but Global removed, the think rate rises further to 99.18%, indicating near-complete collapse to Thinking, while the overall score falls by 1.78 points. These results indicate that sample-level supervision alone is insufficient to prevent single-mode dominance. Global and Sample are complementary: the former mitigates mode imbalance, and the latter improves input-dependent routing.

We hope these additional analyses address your concerns and would greatly appreciate your reconsideration of our work. Thank you again for your valuable feedback.

I appreciate the authors' detailed response and the additional experiments, which have mostly addressed my main concerns. The clarifications and new evidence have helped me better understand the proposed method and its contributions. I suggest incorporating these explanations and experimental findings into the revised manuscript to improve its clarity and completeness. Therefore, I will raise my rating.

Thank you for your careful reading, thoughtful feedback, and recognition of our work. We sincerely appreciate your positive response and willingness to reconsider the rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.