RTWI: Reliable Multimodal Reasoning by Filtering Noisy Thinking

Introduction

With OpenAI's 2025 release of the o3 reasoning model, Thinking with Images (TWI) has become a mainstream paradigm for enhancing multimodal large language model (MLLM) reasoning. TWI operates in two core steps: Cue Mining — examining complex images to extract visual cues, and Answer Reasoning — chaining those cues into logical steps to derive answers. However, existing TWI paradigms ignore noisy thinking caused by erroneous cues and faulty reasoning, leading to a "one step wrong, step by step wrong" error propagation.

Background and Challenges

TWI first generates a textual chain-of-thought to call external tools for visual cues (cue mining), then uses those cues as intermediate steps for textual reasoning (answer reasoning). As shown in Figure 1, both stages can produce noisy thinking. Statistics on real datasets confirm that incorrect answers often trace back to errors in cue mining and answer reasoning. Test-time scaling (TTS) could mitigate this, but existing TTS methods are designed for textual chains only. Visual signals are continuous and unstructured, making uncertainty modeling harder; visual uncertainty does not equal cue relevance. Moreover, no prior work explores how erroneous visual cues propagate and disrupt subsequent textual reasoning.

Main Contributions

First exploration of noisy thinking in TWI, revealing its negative impact from a reasoning reliability perspective.

A unified mechanism modeling visual and textual reliability, uncovering two key properties — reliability correlation and reliability leap — and a plug-and-play robust TTS framework based on them.

The method improves offline answer accuracy and supports online reliable early stopping, significantly boosting inference efficiency.

Two Key Observations on Reasoning Reliability

RTWI establishes a unified text-centric reliability measure for both stages. Instead of directly estimating visual reliability (which is difficult and lacks standards), it leverages the fact that visual cues are obtained via tool-calling instructions; reliable text instructions are a prerequisite for correct cues. The reliability of a reasoning path π at stage s is computed from the entropy of tokens in that stage.

Based on this measure, two core observations emerge (Figure 2):

Reliability Correlation : Higher reliability in both stages strongly correlates with final answer accuracy ("the more confident, the more accurate").

Reliability Leap : Correct visual cues often yield a larger reliability gain from cue mining to answer reasoning ("from hesitant to confident").

Analysis of correct vs. incorrect cases (Figure 3) shows that errors propagate, lowering stage reliability and reliability leap, while correct paths exhibit high reliability in both stages and a pronounced leap.

Method: RTWI Framework

The framework (Figure 4) consists of Dual-Stage Filtering and Credible Voting .

Dual-Stage Filtering

To prevent unreliable paths from biasing the answer, RTWI uses multi-path sampling with a filter-and-retain mechanism. For each stage, a threshold is set at the p -th percentile of reliability scores from a calibration set of paths. Paths below the threshold at either stage are discarded.

Credible Voting

After filtering, instead of simple majority voting, RTWI performs weighted voting where each path's weight combines its reliability and reliability gain, scaled by a temperature coefficient. This gives more influence to paths that are both confident and show a strong reliability leap.

Experimental Results

Experiments use two state-of-the-art TWI models — Qwen3-VL and DeepEyes — across multiple benchmarks: high-resolution datasets, TWI-oriented datasets, multimodal math reasoning, and open-ended VQA. Two settings are evaluated:

Online : Simulates real inference; uses few paths to estimate thresholds, then early-stops unreliable paths within a budget. Focuses on accuracy gain with reduced compute.

Offline : Tests the "filtering" ability — mining correct answers from a given set of paths.

Key Findings

More accurate, faster : RTWI leads in reasoning accuracy while maintaining competitive Token Saving Rate (TSR).

Larger budget, greater advantage : As compute budget increases, RTWI shows steadier performance growth and better performance-per-cost ratio.

Strong scaling law : Across different model sizes, RTWI consistently delivers significant improvements over baselines.

Detailed results are presented in Table 1 (high-resolution online), Figure 5 (test-time scaling analysis), and Figure 6 (scaling law analysis).

Conclusion and Outlook

RTWI introduces the noisy thinking problem in TWI, quantifies its harm via reliability, and provides a practical plug-and-play solution. As agentic LLMs evolve from knowledge bases into autonomous tool-wielding commanders, RTWI's reliable reasoning approach is expected to inspire trustworthy agentic systems in broader scenarios.

Paper: https://arxiv.org/pdf/2602.12916 (ICML 2026). Code: https://github.com/XLearning-SCU/Reliable_TWI

Code example

来源：PaperWeekly
本文
约3200字
，建议阅读
6
分钟
本文介绍了 RTWI 框架，缓解 TWI 噪声思维，提升多模态模型推理可靠性。