PaperlayerView exact revision on arXiv

Source-linked AI summary

Can Textual Reasoning Improve the Performance of MLLMs on Fine-grained Visual Classification?

Jie Zhu, Yiyang Su, Xiaoming Liu

arXiv:2601.06993v2cs.CV

TL;DR

MLLMs struggle with FGVC, and the paper asks why CoT can degrade this perception-heavy task. It systematically studies zero-shot and training settings, identifies reasoning length as the “Cost of Thinking,” and proposes MRN and ReFine-RFT, which achieve state-of-the-art performance across FGVC benchmarks.

  • Problem

    MLLMs struggle to distinguish fine-grained categories, while prior evidence that CoT can harm visual perception leaves why reasoning degrades FGVC insufficiently explained.

  • Method

    The paper evaluates CoT in zero-shot and multiple training regimes, then combines ensemble semantically-aware rewards with MRN in the reasoning-constrained ReFine-RFT framework.

  • Results

    ReFine-RFT achieves state-of-the-art performance across multiple FGVC benchmarks, while longer textual reasoning consistently corresponds to lower visual perception accuracy.

  • Takeaways & Limitations

    Fine-grained visual perception in MLLMs benefits from concise rather than elaborate reasoning and accuracy-centric reward shaping.

  • Takeaways & Limitations

    The study focuses on FGVC, leaves reasoning quality during RFT unexplored, and retains a more time-consuming pipeline than standard SFT.

Abstract

from arXiv · show

Multi-modal large language models (MLLMs) exhibit strong general-purpose capabilities, yet still struggle on Fine-Grained Visual Classification (FGVC), a core perception task that requires subtle visual discrimination and is crucial for many real-world applications. A widely adopted strategy for boosting performance on challenging tasks such as math and coding is Chain-of-Thought (CoT) reasoning. However, several prior works have reported that CoT can actually harm performance on visual perception tasks. These studies, though, examine the issue from relatively narrow angles and leave open why CoT degrades perception-heavy performance. We systematically re-examine the role of CoT in FGVC through the lenses of zero-shot evaluation and multiple training paradigms. Across these settings, we uncover a central paradox: the degradation induced by CoT is largely driven by the reasoning length, in which longer textual reasoning consistently lowers classification accuracy. We term this phenomenon the ``Cost of Thinking''. Building on this finding, we make two key contributions: (1) MRN, a simple and general plug-and-play normalization method for multi-reward optimization that balances heterogeneous reward signals, and (2) ReFine-RFT, a framework that combines ensemble rewards with MRN to constrain reasoning length while providing dense accuracy-oriented feedback. Extensive experiments demonstrate the effectiveness of our findings and the proposed ReFine-RFT, achieving state-of-the-art performance across FGVC benchmarks. Project page: \href{https://refine-rft.github.io/}{ReFine-RFT}.

1. Introduction

The paper investigates why CoT can hurt fine-grained visual perception and finds that longer textual reasoning consistently lowers FGVC accuracy. It introduces MRN and ReFine-RFT to balance reward signals, constrain reasoning length, and improve performance.

  • Motivation: FGVC requires distinguishing subordinate-level categories using subtle visual cues, yet current MLLMs often fail at these distinctions.This limitation is especially relevant in scarce-data and specialized-domain settings.
  • Research question: The paper asks whether textual reasoning helps compensate for MLLMs’ perceptual weaknesses or instead harms fine-grained visual perception.
  • Key finding: Across zero-shot evaluation and RFT, longer textual reasoning consistently correlates with worse visual perception accuracy, a phenomenon termed the “Cost of Thinking”.During RFT, reasoning length shrinks while accuracy improves, producing reasoning collapse.
  • Method: ReFine-RFT combines ensemble, semantically aware rewards with MRN to provide dense accuracy-oriented feedback while explicitly constraining reasoning length.MRN independently normalizes heterogeneous rewards and reduces reward-signal variance.
  • Related work: Prior work established that CoT may degrade visual perception, while this study re-evaluates the issue across zero-shot and multiple training regimes and identifies reasoning length as the key factor.

3. Cost of Thinking in FGVC

Across FGVC experiments, longer textual reasoning consistently lowers classification accuracy, while RFT learns to shorten reasoning and favor concise responses. These findings motivate ReFine-RFT, which combines ensemble rewards with MRN to constrain reasoning length and stabilize optimization.

  • Experiment Setup: FGVC evaluation uses four benchmarks in a 4-shot open-ended question-answering setting.The datasets are FGVC-Aircraft, Stanford-Cars, Flowers-102, and Oxford-Pets.
  • Experiment Setup: Answer-only and CoT prompts are evaluated across Qwen2/2.5-VL, InternVL, and R1-OneVision models.RFT uses Qwen2-VL-2B as the base model.
  • Experiment Setup: The format reward enforces structured <think>...</think> and <answer>...</answer> outputs, while the classification reward scores whether the predicted label matches the ground truth.The format and classification rewards are binary signals.
  • Experiment Setup: A thinking-length reward assigns a binary score when reasoning falls within predefined bounds and otherwise provides no length-based reward.The reward uses the extracted reasoning content length and can be zero when formatting is invalid.
  • Zero-Shot Results: 3–6% average accuracy drops occur when non-reasoning models switch from Answer-only to CoT prompts across FGVC datasets.The paper reports that CoT can introduce useless or hallucinatory explanations rather than improve decision quality.
  • RFT Results: During RFT, average reasoning length decreases across all datasets and stabilizes below the zero-shot level, a pattern termed reasoning collapse.The authors describe this as RFT suppressing reasoning steps that do not contribute to reward maximization.
  • Reasoning-Length Analysis: Enforcing longer reasoning sequences lowers classification accuracy across FGVC datasets, whereas shorter traces yield higher accuracy.The study varies reasoning-length limits from [0, 20] to [60, 80] and notes that reasoning quality is unsupervised during RFT.
  • Training-Paradigm Results: SFT-AO consistently outperforms SFT-CoT even when GPT-4o supplies high-quality CoT data, supporting a length-based Cost of Thinking across training and inference.The comparison indicates that the degradation is not simply attributed to CoT quality.

4. Methods

ReFine-RFT combines ensemble rewards with Multi-Reward Normalization to provide dense accuracy feedback, constrain reasoning length, and stabilize optimization across heterogeneous rewards.

  • Ensemble Reward: ReFine-RFT combines format, accuracy, thinking-length, MLLM-based accuracy, and embedding-similarity rewards for richer, accuracy-centric feedback.The ensemble addresses binary accuracy supervision and includes semantically aware signals.
  • Ensemble Reward: The embedding-similarity reward uses cosine similarity between predicted and ground-truth text embeddings to provide differentiable semantic supervision.This encourages semantically close answers even when their lexical forms differ.
  • GRPO: GRPO samples diverse responses, computes rewards for each, normalizes them using group statistics, and aggregates the resulting advantages.Group normalization encourages responses with higher reward.
  • Multi-Reward Normalization (MRN): MRN independently normalizes heterogeneous reward signals before aggregation, addressing differing convergence rates and preventing rapidly saturating rewards from dominating optimization.Figure 5 highlights differences in convergence speed, value range, and saturation point among rewards.

5. Experiments

Experiments evaluate ReFine-RFT across FGVC datasets, training regimes, reward designs, and reasoning controls. The results consistently favor RFT, MRN, ensemble rewards, and controlled reasoning length.

  • Results of ReFine-RFT: RFT yields substantial accuracy improvements over SFT across all FGVC benchmarks by optimizing accuracy-centric objectives.The paper reports more stable and discriminative predictions, especially in few-shot FGVC.
  • Results of ReFine-RFT: ReFine-RFT with a 2B backbone and 4-shot training significantly surpasses Finedefics-8B trained on full FGVC datasets.The comparison underscores the reported efficiency and scalability of the approach.
  • Results of ReFine-RFT: ReFine-RFT achieves the best results with consistent gains across FGVC datasets over SFT and Visual-RFT baselines.Table 2 compares SFT variants and Visual-RFT, with parentheses denoting relative improvements over Visual-RFT (LoRA).
  • Reasoning Control: Controlling reasoning length has a larger performance impact than prompt style, while ReFine-RFT-CoT is slightly better than ReFine-RFT-AO.No-Thinking-RFT and Visual-RFT achieve similar performance despite using answer-only and CoT prompts, respectively.
  • Ablation Studies: MRN improves accuracy across three FGVC datasets by +1.1%/+2.1%/+0.4% with full fine-tuning and +0.7%/+1.5%/+0.6% with LoRA.The gains are more pronounced with larger training capacity and remain effective in parameter-efficient training.
  • Ablation Studies: Combining multiple reward functions consistently improves performance, with all reward functions jointly producing the best overall result.The reward components contribute complementary information to robust learning objectives.
  • Ablation Studies: Larger LoRA capacity progressively improves accuracy, and r=64, α=128 surpasses the Fully-FT baseline across all datasets.The smaller r=16, α=32 configuration causes a notable performance drop.
  • Ablation Studies: MRN achieves higher rewards and lower reward variance than baseline GRPO throughout training, indicating more stable optimization.The comparison is reported on Aircrafts-102.

6. Conclusion

The paper identifies a Cost of Thinking in FGVC: excessive textual reasoning degrades MLLM perception performance. ReFine-RFT addresses this through reasoning constraints, ensemble rewards, and MRN.

  • Conclusion: The Cost of Thinking describes degradation from excessive textual reasoning across both inference and training perspectives in FGVC.The study concludes that perception-centric tasks benefit more from concise reasoning.
  • Conclusion: ReFine-RFT combines reasoning constraints, ensemble semantically aware rewards, and MRN to enhance visual perception.The paper reports state-of-the-art performance across FGVC benchmarks.

7. GRPO Algorithm

GRPO samples multiple responses, evaluates their rewards, and normalizes rewards within each group to guide policy updates toward preferred answers. The update also uses clipping and a KL penalty relative to a reference model.

  • GRPO samples G diverse responses from the current policy and obtains a reward for each response.
  • Group normalization uses the group reward mean and standard deviation to assess relative response quality.
  • The policy update encourages responses with higher normalized rewards while applying clipping and a KL penalty controlled by ε and β.

8. Additional Implementation Details

The experiments use fixed data, prompting, model, training, and evaluation procedures. These details specify how answer-only and CoT variants are trained, rewarded, and assessed.

  • Training and evaluation use the provided 4-shot data, with dataset statistics reported in Table 6.
  • Answer-only prompts are used for SFT, whereas CoT prompts are used for CoT-SFT and RFT training.
  • Qwen2-VL-7B serves as the MLLM-based accuracy reward model, deployed with LMDeploy and mixed-precision TurboMind inference.
  • GPT-4o-2024-08-06 generates CoT data from images, prompts, and ground-truth labels to produce reasoning leading to correct answers.
  • Experiments use fixed seeds, BF16, LoRA with rank γ = 64 and α = 128, a 200-step maximum, and a 256-token completion cap.
  • CoT predictions are evaluated by normalized substring matching between extracted answer tags and ground truth, while answer-only outputs are compared directly.

9. Additional Experimental Results

Additional experiments compare prompt types, discriminative baselines, reward combinations, and reasoning lengths across datasets. They consistently support concise reasoning and show that ReFine-RFT remains competitive across several FGVC settings.

  • Prompt-type comparisons: Answer-only and CoT RFT prompts achieve almost identical accuracy on Aircrafts-102 and Cars-196, indicating little benefit from explicit long traces.
  • Discriminative comparisons: ReFine-RFT outperforms CLIPLP by +10.4% on Cars and +19.8% on Aircrafts, while linear-probed CLIP remains more competitive on Flowers and Pets.
  • Reward analysis: Combining all rewards yields the best performance, while classification, embedding, and MLLM rewards show consistent positive trends.
  • Reward analysis: Classification, embedding, and MLLM rewards are correlated yet distinct, providing complementary learning signals rather than redundant supervision.
  • Reward analysis: Format and thinking-length rewards correlate more weakly with task-related rewards and primarily regularize output structure and response behavior.
  • Thinking length: ReFine-RFT is most concise across datasets, while SFT-CoT is longest, and the Cars and Aircrafts gaps exceed twofold or threefold.

10. Potential Reasons of CoT Degradation on Visual Tasks.

The paper attributes CoT degradation to interactions between reasoning length, visual processing, and noisy intermediate descriptions. Longer chains can divert attention from image evidence and accumulate errors in fine-grained decisions.

  • Long CoT sequences may compete with visual processing for finite attention and context, increasing reliance on linguistic history over image embeddings.
  • Answer-only RFT reaches almost identical accuracy to explicit CoT RFT, while ReFine-RFT produces the shortest reasoning across datasets.
  • Additional reasoning steps create more opportunities for hallucinated details, spurious correlations, and incorrect intermediate visual descriptions.
  • In FGVC, an incorrect local description can be reinforced by subsequent reasoning instead of being corrected through renewed image inspection.

11. Potential Social Impact

The work highlights potential benefits of fine-grained visual recognition for high-stakes and expert applications, while pairing accurate predictions with interpretable reasoning steps.

  • Fine-grained categorization could benefit biodiversity monitoring, medical diagnostics, industrial inspection, and scientific research.
  • Interpretable reasoning steps alongside accurate predictions may promote transparency and trustworthiness in these applications.

12. Limitation

ReFine-RFT remains more time-consuming than standard SFT, leaves reasoning-quality effects unexplored, and has been studied in detail only for FGVC.

  • Despite indirectly improving training efficiency by suppressing excessive thinking length, ReFine-RFT remains more time-consuming than standard SFT.The additional cost comes from rollout sampling and RFT optimization.
  • The analysis focuses mainly on thinking length and comparisons among SFT-AO, SFT-CoT, and RFT variants, leaving reasoning quality during RFT unexplored.
  • The detailed study is limited to fine-grained visual classification, motivating extension to object detection, visual grounding, and open-ended vision–language reasoning.
Loading 2601.06993v2…