Pith. sign in

REVIEW 3 major objections 25 references

Multimodal models can learn to choose when to reason and when to answer directly, cutting unnecessary thinking while keeping accuracy on mixed tasks.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-10 05:06 UTC pith:KLGNQ6LA

load-bearing objection Solid systems paper: toolized think/direct under GRPO plus dual-level regulation gives a real accuracy–efficiency win on mixed multimodal tasks, with the main soft spot being threshold-sensitive route labels rather than a broken claim. the 3 major comments →

arxiv 2607.08572 v1 pith:KLGNQ6LA submitted 2026-07-09 cs.CV

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

Yiyang Fang , Pei Fu , Jinjie Li , Jian Liang , Wenke Huang , Ruijie Luo , Shaojie Zhang , Jian Luan
show 2 more authors
Yi R. Fung Mang Ye
This is my paper
classification cs.CV
keywords adaptive reasoningmultimodal large language modelsGRPOthinking modedirect modemode collapsereasoning efficiencytool invocation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Multimodal language models are usually trained to think step by step before every answer. That fixed habit wastes tokens and time on easy questions, while always answering immediately hurts hard ones. This paper argues that the real problem in mixed multimodal training is not only accuracy but learning a stable policy for when explicit reasoning helps. Switch-Reasoner turns thinking into a virtual tool the model can call or skip, then trains that choice with group-relative reinforcement learning. A dual-level controller keeps the two modes from collapsing into always-think or always-direct behavior: one signal balances overall mode use across batches, and another labels each sample by whether forced thinking actually improves success over a forced direct answer. On eleven multimodal tasks the method matches or beats always-thinking training at roughly half the thinking rate on the smaller model, and at under forty percent on the larger one.

Core claim

In heterogeneous multimodal mixtures, treating explicit reasoning as an optional tool call and regulating that choice at both batch and sample levels yields a better accuracy–efficiency trade-off than fixed always-think or always-direct GRPO: the model learns instance-adaptive switching instead of collapsing to one mode.

What carries the argument

Thinking-as-Tool with dual-level mode regulation: the model first emits either a direct answer or a require_think tool call; training then applies a delayed global balance signal on mode usage and accuracy, plus sample-level route rewards from counterfactual success-rate gaps between forced thinking and forced direct answering.

Load-bearing premise

That a few forced counterfactual rollouts per sample, plus hand-set thresholds, correctly tell the model which inputs truly need thinking rather than reflecting prompt format or short-run reward noise.

What would settle it

Train the same base models on the same mixture with the dual-level controller ablated or with K=0 counterfactual labels, then check whether overall score falls and thinking rate collapses toward 0% or 100% on held-out heterogeneous tasks such as Geometry3k versus ChartQA.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Fixed Think-then-Answer post-training is not required for strong multimodal accuracy; selective mode choice can match or exceed it at lower thinking cost.
  • Mode collapse under GRPO in multitask mixtures can be reduced by combining batch-level usage balance with per-sample counterfactual route labels.
  • Reasoning-heavy tasks (geometry, olympiad-style items) should receive higher thinking rates than recognition or chart lookup tasks under the same policy.
  • The tool-call interface can later host richer reasoning tools beyond a binary think/direct switch.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If forced-route success gaps are noisy, safer designs might replace fixed thresholds with learned value estimates of each mode under a shared reward.
  • The same switchable interface could gate other expensive actions—vision tools, search, or multi-hop planning—using the same global-plus-sample regulation pattern.
  • Deployment systems could expose thinking rate as a controllable knob by reweighting the global balance signal without retraining the whole policy.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 0 minor

Summary. Switch-Reasoner addresses fixed Think-then-Answer inefficiency in heterogeneous multimodal RL by casting explicit reasoning as a virtual tool call under a switchable prompt, so the policy can choose Direct Mode or Thinking Mode before generating long traces. To prevent GRPO mode collapse, it adds dual-level regulation: a delayed global controller B that rebalances mode usage from prior-batch accuracy and count statistics (Eqs. 3–6), and sample-level preferred-route labels from forced Direct/Thinking counterfactual success rates (Eqs. 7–11), combined into the final reward r*_i (Eq. 17). On Qwen3-VL-4B/8B trained over 11 multimodal tasks, the method reports better accuracy–efficiency trade-offs than always-think and always-direct GRPO (Table 1), with selective dataset-level think rates (Fig. 4), training dynamics (Fig. 3), case studies, and hyperparameter ablations (Table 4).

Significance. If the result holds, the paper offers a practical, GRPO-native way to learn when to think in multitask MLLM mixtures without multi-stage curricula. The Thinking-as-Tool interface is simple and reusable; dual-level regulation is a concrete response to mode collapse that many adaptive-reasoning papers leave to schedules or cold starts. Strengths include a clear multitask evaluation (11 tasks), explicit think-rate reporting, training-dynamics and dataset-level visualizations, case studies, a full algorithm/prompt appendix, and hyperparameter ablations. Gains over always-direct are large; over always-think they are modest but paired with large think-rate reductions (e.g., 4B: 71.60 at 51.53% vs 70.66 at 100%). Code is linked. The contribution is incremental relative to concurrent adaptive-reasoning work, but the multitask GRPO stabilization angle is useful for efficient multimodal post-training.

major comments (3)
  1. §3.4–3.5, Eqs. 7–11 and 17: the central claim that the model learns instance-level reasoning demand rests on preferred-route labels ci derived from forced-route success rates with default K=4. Appendix E / Table 4 shows the accuracy–efficiency trade-off is highly sensitive to the safe-direct criterion (loose: 4B overall 68.32 at 13.28% think rate; strict: 70.77 at 82.48%). Without evidence that forced-prompt gaps are not prompt-format artifacts or small-K noise—e.g., human agreement on must-think/safe-direct, held-out label stability, or comparison to difficulty proxies independent of the same policy—the dual-level reward may partly manufacture the selective policy rather than discover genuine demand. A targeted validation or stronger baseline that removes forced counterfactuals while keeping the switchable interface would make the claim load-bearing rather than heuristic-dependent.
  2. Table 1 and §4.2: primary comparisons are Vanilla, GRPO-Thinking, GRPO-Direct, and an internal ablation without sample-level correction. Related adaptive-reasoning methods cited in §2.2 (constrained optimization, cold-start, staged RL, difficulty-aware curricula) are not reimplemented under the same GRPO/multitask mixture. The reported trade-off is therefore only relative to fixed always-think/always-direct GRPO. At least one strong external adaptive baseline (or a clear statement that none is feasible under identical data/compute, with a proxy) is needed to support the claim of a better accuracy–efficiency solution for heterogeneous multimodal mixtures rather than only for the fixed-mode GRPO endpoints.
  3. §3.3, Eqs. 5–6 and Algorithm 1: global control B is a delayed, thresholded heuristic (τ_low/high, δ, π_low/high, Bg/Bf) applied as a multiplicative penalty. The paper does not report mode-usage trajectories under ablated global control alone, nor sensitivity of collapse rates to these thresholds comparable to the sample-level ablations in Table 4. Given that mode collapse is a stated motivation (Fig. 1, Introduction), the contribution of the global controller versus sample-level route rewards should be isolated more clearly—e.g., Switch-Reasoner without B, or with B only—so that dual-level regulation is not credited as a package when one level may dominate.

Circularity Check

0 steps flagged

No significant circularity: empirical GRPO method with online self-labels and external multitask benchmarks, not a derivation that restates inputs as predictions.

full rationale

Switch-Reasoner is an empirical RL post-training paper, not a first-principles derivation. Its central claim is a measured accuracy–efficiency trade-off on 11 external multimodal benchmarks (Table 1; Figures 3–5), comparing against always-think and always-direct GRPO baselines under the same GRPO setup. The dual-level mechanism uses (i) a one-batch-delayed global controller B from prior-batch usage/accuracy statistics (Eqs. 3–6, 13) and (ii) sample-level preferred-route labels ci from forced-route success rates under the current policy (Eqs. 7–11, 14–16). Those labels are training signals, not claimed predictions; the paper explicitly delays B to avoid same-batch circular dependence between rewards and the statistics computed from them. Using the policy’s own counterfactual rollouts for route supervision is standard online RL self-training and does not make the reported validation scores true by construction—the method can and does underperform under ablated thresholds (Appendix E). There is no self-definitional identity, no fitted constant renamed as a prediction, no load-bearing uniqueness theorem imported from the authors, and no ansatz smuggled in via self-citation. Concerns that forced-prompt gaps may be imperfect proxies for true reasoning demand are validity/correctness risks, not circular reductions of the result to its inputs.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 2 invented entities

The central claim is empirical and rests on GRPO/RLHF practice, a binary mode interface, accuracy/format rewards, and many hand-set control thresholds. No new physical entities; invented constructs are algorithmic (virtual think tool, dual controllers). Free parameters are the routing thresholds and route-reward weights that materially shape think rate and score (Table 4).

free parameters (4)
  • Global mode thresholds (τ_low, τ_high, δ, π_low, π_high) and signal magnitudes (B_g, B_f)
    Hand-chosen bands that decide when to favor Direct vs Thinking and when extreme-imbalance safeguards fire; they directly regulate mode collapse behavior.
  • Sample-level route thresholds (δ_think, τ_direct, δ_safe) and weights (w_think, w_direct)
    Define must-think/safe-direct/uncertain labels and route-reward strength; ablations show safe-direct settings strongly change accuracy–efficiency trade-off.
  • Counterfactual rollout count K and GRPO group size G
    K controls reliability of empirical success rates used as supervision; default K=4 is a design choice with measurable sensitivity.
  • Base reward mix α and global penalty factor η
    Trade off accuracy vs format compliance and strength of global mode penalties in the final reward used for advantages.
axioms (5)
  • domain assumption Group Relative Policy Optimization with group-normalized advantages is a valid post-training objective for multimodal reasoning policies.
    Section 3.1 treats GRPO as the optimization backbone without re-deriving its optimality properties.
  • ad hoc to paper Whether-to-think can be adequately represented as a binary discrete action (tool call vs direct answer) before generation.
    Thinking-as-Tool (Section 3.2) assumes a two-mode interface captures the efficiency problem; continuous budgets or multi-tool policies are left future work.
  • domain assumption Task accuracy (plus format compliance) is a sufficient primary reward signal for learning useful mode selection.
    Reward design (Section 3.5) centers r_acc and r_fmt; latency/token cost is optimized only indirectly via mode choice.
  • ad hoc to paper Forced-route success-rate gaps estimate the true sample-specific benefit of explicit reasoning.
    Equations 7–11 treat Δ_i from K forced rollouts as counterfactual route supervision for preferred-mode labels.
  • domain assumption Format-valid trajectories are the right population for estimating mode usage and accuracy statistics.
    Section 3.3 discards malformed outputs when computing n_think/n_direct and accuracies for the global controller.
invented entities (2)
  • Virtual thinking tool T_think / require_think tool call no independent evidence
    purpose: Make mode selection an explicit, optimizable action under a switchable prompt protocol.
    Introduced in Section 3.2 as a virtual tool, not an external API with independent existence outside the training interface.
  • Dual-level mode regulation (global delayed controller B + sample-level route categories) no independent evidence
    purpose: Prevent always-think/always-direct collapse and provide instance-adaptive route supervision during GRPO.
    Core algorithmic construct of the paper (Sections 3.3–3.5); validated only within the reported experiments.

pith-pipeline@v1.1.0-grok45 · 23512 in / 3376 out tokens · 37260 ms · 2026-07-10T05:06:45.718472+00:00 · methodology

0 comments
read the original 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.

Figures

Figures reproduced from arXiv: 2607.08572 by Jian Liang, Jian Luan, Jinjie Li, Mang Ye, Pei Fu, Ruijie Luo, Shaojie Zhang, Wenke Huang, Yi R. Fung, Yiyang Fang.

Figure 1
Figure 1. Figure 1: Motivation for selective reasoning in heterogeneous multimodal tasks. Left: Tasks differ in their need for explicit reasoning. Top right: Accuracy-only training can collapse to one mode. Bottom right: Heterogeneous mixtures make stable Thinking/Direct Mode selection difficult. reducing output length, they do not address the preceding decision of whether explicit reasoning is needed at all. A more closely r… view at source ↗
Figure 2
Figure 2. Figure 2: Overall framework of Switch-Reasoner. The model first selects between Direct Mode and Thinking Mode through a switchable prompt. Tool invocation triggers explicit reasoning, while direct answers bypass it. During training, a global controller regulates mode balance, and a sample￾level optimizer provides counterfactual route supervision. where β is the KL regularization hyperparameter, and fϵ(x, y) = min(xy… view at source ↗
Figure 3
Figure 3. Figure 3: Training dynamics of validation performance and thinking rate. Refer to Section [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Dataset-level score and thinking rate of the 4B model. See Section [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Case studies on different datasets with the 4B model. Refer to Section [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Broader case studies (1) 18 [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Broader case studies (2) 19 [PITH_FULL_IMAGE:figures/full_fig_p019_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

25 extracted references · 25 canonical work pages · 21 internal anchors

  1. [1]

    Qwen3-VL Technical Report

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-vl technical report.arXiv preprint arXiv:2511.21631,

    work page internal anchor Pith review Pith/arXiv arXiv

  2. [2]

    Ares: Multimodal adaptive reasoning via difficulty-aware token-level entropy shaping.arXiv preprint arXiv:2510.08457,

    Shuang Chen, Yue Guo, Yimeng Ye, Shijue Huang, Wenbo Hu, Haoxi Li, Manyuan Zhang, Jiayu Chen, Song Guo, and Nanyun Peng. Ares: Multimodal adaptive reasoning via difficulty-aware token-level entropy shaping.arXiv preprint arXiv:2510.08457,

    work page arXiv

  3. [3]

    Video-R1: Reinforcing Video Reasoning in MLLMs

    Gongfan Fang, Xinyin Ma, and Xinchao Wang. Thinkless: Llm learns when to think. InNeurIPS, volume 38, pp. 151268–151295, 2026a. Yiyang Fang, Wenke Huang, Pei Fu, Yihao Yang, Kehua Su, Zhenbo Luo, Jian Luan, and Mang Ye. Emo-r3: Reflective reinforcement learning for emotional reasoning in multimodal large language models. InCVPR, pp. 745–755, 2026b. Kaituo...

    work page internal anchor Pith review Pith/arXiv arXiv

  4. [4]

    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948,

    work page internal anchor Pith review Pith/arXiv arXiv

  5. [5]

    ThinkPrune: Pruning Long Chain-of-Thought of LLMs via Reinforcement Learning

    Bairu Hou, Yang Zhang, Jiabao Ji, Yujian Liu, Kaizhi Qian, Jacob Andreas, and Shiyu Chang. Thinkprune: Pruning long chain-of-thought of llms via reinforcement learning.arXiv preprint arXiv:2504.01296,

    work page internal anchor Pith review Pith/arXiv arXiv

  6. [6]

    Boosting MLLM Reasoning with Text-Debiased Hint-GRPO

    Qihan Huang, Weilong Dai, Jinlong Liu, Wanggui He, Hao Jiang, Mingli Song, Jingyuan Chen, Chang Yao, and Jie Song. Boosting mllm reasoning with text-debiased hint-grpo.arXiv preprint arXiv:2503.23905, 2025a. Shijue Huang, Hongru Wang, Wanjun Zhong, Zhaochen Su, Jiazhan Feng, Bowen Cao, and Yi R Fung. Adactrl: Towards adaptive and controllable reasoning vi...

    work page internal anchor Pith review Pith/arXiv arXiv

  7. [7]

    LLaVA-OneVision: Easy Visual Task Transfer

    Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer.arXiv preprint arXiv:2408.03326, 2024a. 10 Preprint Xinhao Li, Ziang Yan, Desen Meng, Lu Dong, Xiangyu Zeng, Yinan He, Yali Wang, Yu Qiao, Yi Wang, and Limin Wang. Videochat-r1: Enhancing spatio...

    work page internal anchor Pith review Pith/arXiv arXiv

  8. [8]

    Reinforcement Learning Meets Large Language Models: A Survey of Advancements and Applications Across the LLM Lifecycle

    Keliang Liu, Dingkang Yang, Ziyun Qian, Weijie Yin, Yuchi Wang, Hongsheng Li, Jun Liu, Peng Zhai, Yang Liu, and Lihua Zhang. Reinforcement learning meets large language models: A survey of advancements and applications across the llm lifecycle.arXiv preprint arXiv:2509.16679, 2025a. Shuming Liu, Mingchen Zhuge, Changsheng Zhao, Jun Chen, Lemeng Wu, Zechun...

    work page internal anchor Pith review Pith/arXiv arXiv

  9. [9]

    Visual-RFT: Visual Reinforcement Fine-Tuning

    Ziyu Liu, Zeyi Sun, Yuhang Zang, Xiaoyi Dong, Yuhang Cao, Haodong Duan, Dahua Lin, and Jiaqi Wang. Visual-rft: Visual reinforcement fine-tuning.arXiv preprint arXiv:2503.01785, 2025b. Chenwei Lou, Zewei Sun, Xinnian Liang, Meng Qu, Wei Shen, Wenqi Wang, Yuntao Li, Qingping Yang, and Shuangzhi Wu. Adacot: Pareto-optimal adaptive chain-of-thought triggering...

    work page internal anchor Pith review Pith/arXiv arXiv

  10. [10]

    IconQA: A New Benchmark for Abstract Diagram Understanding and Visual Language Reasoning

    Pan Lu, Ran Gong, Shibiao Jiang, Liang Qiu, Siyuan Huang, Xiaodan Liang, and Song-Chun Zhu. Inter-gps: Interpretable geometry problem solving with formal language and symbolic reasoning. InACL, pp. 6774–6786, 2021a. Pan Lu, Liang Qiu, Jiaqi Chen, Tony Xia, Yizhou Zhao, Wei Zhang, Zhou Yu, Xiaodan Liang, and Song-Chun Zhu. Iconqa: A new benchmark for abstr...

    work page internal anchor Pith review Pith/arXiv arXiv

  11. [11]

    TERMINATOR: Learning Optimal Exit Points for Early Stopping in Chain-of-Thought Reasoning

    Alliot Nagle, Jakhongir Saydaliev, Dhia Garbaya, Michael Gastpar, Ashok Vardhan Makkuva, and Hyeji Kim. Terminator: Learning optimal exit points for early stopping in chain-of-thought reasoning.arXiv preprint arXiv:2603.12529,

    work page internal anchor Pith review Pith/arXiv arXiv

  12. [12]

    Safegrpo: Self-rewarded multimodal safety alignment via rule-governed policy optimization.arXiv preprint arXiv:2511.12982,

    11 Preprint Xuankun Rong, Wenke Huang, Tingfeng Wang, Daiguo Zhou, Bo Du, and Mang Ye. Safegrpo: Self-rewarded multimodal safety alignment via rule-governed policy optimization.arXiv preprint arXiv:2511.12982,

    work page arXiv

  13. [13]

    Proximal Policy Optimization Algorithms

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347,

    work page internal anchor Pith review Pith/arXiv arXiv

  14. [14]

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathemat- ical reasoning in open language models.arXiv preprint arXiv:2402.03300,

    work page internal anchor Pith review Pith/arXiv arXiv

  15. [15]

    Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution

    Chaoyang Wang, Kaituo Feng, Dongyang Chen, Zhongyu Wang, Zhixun Li, Sicheng Gao, Meng Meng, Xu Zhou, Manyuan Zhang, Yuzhang Shang, et al. Adatooler-v: Adaptive tool-use for images and videos. InACL Findings, pp. 18058–18072, 2026a. Jiaqi Wang, Kevin Qinghong Lin, James Cheng, and Mike Zheng Shou. Think or not? selective reasoning via reinforcement learnin...

    work page internal anchor Pith review Pith/arXiv arXiv

  16. [16]

    AdaptR1: Reinforcement Learning Based Adaptive Interleaved Thinking in Multi-hop Question Answering

    Yuxin Wang, Jiahao Lu, Qifeng Wu, Shicheng Fang, Chuanyuan Tan, Yining Zheng, Xuanjing Huang, and Xipeng Qiu. Adaptr1: Reinforcement learning based adaptive interleaved thinking in multi-hop question answering.arXiv preprint arXiv:2605.31062, 2026c. Shu Wu, Chenxing Li, Wenfu Wang, Hao Zhang, Hualei Wang, Meng Yu, and Dong Yu. Audio- thinker: Guiding larg...

    work page internal anchor Pith review Pith/arXiv arXiv

  17. [17]

    When More is Less: Understanding Chain-of-Thought Length in LLMs

    Yuyang Wu, Yifei Wang, Ziyu Ye, Tianqi Du, Stefanie Jegelka, and Yisen Wang. When more is less: Understanding chain-of-thought length in llms.arXiv preprint arXiv:2502.07266,

    work page internal anchor Pith review Pith/arXiv arXiv

  18. [18]

    Dynamic early exit in reasoning models.arXiv preprint arXiv:2504.15895, 2025a

    Chenxu Yang, Qingyi Si, Yongjie Duan, Zheliang Zhu, Chenyu Zhu, Qiaowei Li, Minghui Chen, Zheng Lin, and Weiping Wang. Dynamic early exit in reasoning models.arXiv preprint arXiv:2504.15895, 2025a. Dongchao Yang, Songxiang Liu, Disong Wang, Yuanyuan Wang, Guanglu Wan, and Helen Meng. Omni-autothink: Adaptive multimodal reasoning via reinforcement learning...

    work page arXiv

  19. [19]

    R-4B: Incentivizing General-Purpose Auto-Thinking Capability in MLLMs via Bi-Mode Annealing and Reinforce Learning

    Qi Yang, Bolin Ni, Shiming Xiang, Han Hu, Houwen Peng, and Jie Jiang. R-4b: Incentivizing general-purpose auto-thinking capability in mllms via bi-mode annealing and reinforce learning. arXiv preprint arXiv:2508.21113, 2025c. Senqiao Yang, Junyi Li, Xin Lai, Jinming Wu, Wei Li, Zejun MA, Bei Yu, Hengshuang Zhao, and Jiaya Jia. Visionthink: Smart and effic...

    work page internal anchor Pith review Pith/arXiv arXiv

  20. [20]

    R1-ShareVL: Incentivizing Reasoning Capability of Multimodal Large Language Models via Share-GRPO

    Huanjin Yao, Qixiang Yin, Jingyi Zhang, Min Yang, Yibo Wang, Wenhao Wu, Fei Su, Li Shen, Minghui Qiu, Dacheng Tao, et al. R1-sharevl: Incentivizing reasoning capability of multimodal large language models via share-grpo.arXiv preprint arXiv:2505.16673,

    work page internal anchor Pith review Pith/arXiv arXiv

  21. [21]

    A Survey of Safety on Large Vision-Language Models: Attacks, Defenses and Evaluations

    Mang Ye, Xuankun Rong, Wenke Huang, Bo Du, Nenghai Yu, and Dacheng Tao. A survey of safety on large vision-language models: Attacks, defenses and evaluations.arXiv preprint arXiv:2502.14881,

    work page internal anchor Pith review Pith/arXiv arXiv

  22. [22]

    R1-VL: Learning to Reason with Multimodal Large Language Models via Step-wise Group Relative Policy Optimization

    Jiajie Zhang, Nianyi Lin, Lei Hou, Ling Feng, and Juanzi Li. Adaptthink: Reasoning models can learn when to think. InEMNLP, pp. 3716–3730, 2025a. Jingyi Zhang, Jiaxing Huang, Huanjin Yao, Shunyu Liu, Xikun Zhang, Shijian Lu, and Dacheng Tao. R1-vl: Learning to reason with multimodal large language models via step-wise group relative policy optimization.ar...

    work page internal anchor Pith review Pith/arXiv arXiv

  23. [23]

    Reinforced MLLM: A Survey on RL-Based Reasoning in Multimodal Large Language Models

    Guanghao Zhou, Panjia Qiu, Cen Chen, Jie Wang, Zheming Yang, Jian Xu, and Minghui Qiu. Reinforced mllm: A survey on rl-based reasoning in multimodal large language models.arXiv preprint arXiv:2504.21277,

    work page internal anchor Pith review Pith/arXiv arXiv

  24. [24]

    Think in Blocks: Adaptive Reasoning from Direct Response to Deep Reasoning

    Yekun Zhu, Guang Chen, and Chengjun Mao. Think in blocks: Adaptive reasoning from direct response to deep reasoning.arXiv preprint arXiv:2508.15507,

    work page internal anchor Pith review Pith/arXiv arXiv

  25. [25]

    Mathe- matics includes Geometry3k and MathVista; Chart/Doc comprises ChartQA and DocVQA; Ground- ing consists of RefAdv; and the remaining benchmarks are grouped under General

    Computes direct i ands think i Compute∆ i =s think i −s direct i if∆ i ≥δ think then ci ←must-think else ifs direct i ≥τ direct and∆ i ≤ −δ safe then ci ←safe-direct else ci ←uncertain end if end for Reward computation Computer base i fromr acc i andr fmt i Apply previous-batch global signalBto obtainr global i Compute route rewardr route i according toc ...

    work page 2024