Research map / Vision-language and multimodal / Multimodal LLMs

Turning Drift into Constraint: Robust Reasoning Alignment in Non-Stationary Multi-Stream Environments

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Autonomous Preference Optimization (APO), which bootstraps a Qwen2.5-VL-7B target from multiple MLLM reasoning trajectories and then uses source-model disagreements as negatives in a multi-negative Plackett-Luce objective, evaluated on chest X-ray classification and report generation, and releases CXR-MAX.

Problem

Aligning a target model's reasoning with multiple non-stationary source MLLMs whose reasoning distributions drift, which transmits biases and hallucinations to the target.

Approach

Treats inter-model divergent reasoning trajectories as dynamic negative constraints: a consensus trajectory synthesized in-context is the positive, and all source trajectories are negatives in a multi-negative Plackett-Luce (DPO-style) preference loss, applied after supervised bootstrapping.

Key results

• On MS-CXR-T multi-label classification, the 7B target reaches average accuracy 0.78, versus 0.69 for the best listed source model (GLM-4.5V) and 0.61-0.69 for full-data specialist baselines, with CoCa-CXR second at 0.69.
• Ablation: adding multiple teachers alone gives 0.68 average vs 0.66 for SPD only; adding APO raises it to 0.78.
• Zero-shot AUC of 0.85 on Open-I and 0.83 on ChestXray14, above CARZero and CPO (0.84/0.82 each).

Surprising findings

• Naively adding multiple teachers yields only marginal overall gains and degrades several pathologies (Con., PE, Pna.), per the ablation.
• The 7B target outperforms every listed proprietary source MLLM on average accuracy despite being much smaller.

Limitations

• Evaluated only on chest X-ray tasks; generalization to other domains is not shown.
• Conservative consensus behavior trades recall for safety on some classes (e.g., edema).
• Claimed superiority over full-data baselines rests on a single narrow classification benchmark with no reported variance.

Experimental setup

Datasets: MIMIC-CXR MS-CXR-T Open-I ChestXray14 ChestXpert ChestXDet10

Benchmarks: MS-CXR-T MIMIC-CXR Open-I ChestXray14 ChestXpert ChestXDet10

Metrics: top-1 accuracy BLEU-1/2/3/4 ROUGE-L METEOR AUC

Baselines: CheXRelNet BioViL CTrans BioViL-T Med-ST TempA-VLP CoCa-CXR METransformer DCL R2GenGPT PromptMRG BtspLLM

Builds on: DPO Plackett-Luce model Bradley-Terry model MIMIC-CXR concept drift theory

Closest papers by meaning

• Walking the Tightrope: Autonomous Disentangling Beneficial and Detrimental Drifts in Non-Stationary Custom-Tuning (NeurIPS 2025 · similarity 0.90)
• RFEval: Benchmarking Reasoning Faithfulness under Counterfactual Reasoning Intervention in Large Reasoning Models (ICLR 2026 · similarity 0.88)
• EcoAlign: An Economically Rational Framework for Efficient LVLM Alignment (CVPR 2026 · similarity 0.88)
• Bias-Variance Optimized Preference Optimization for Large Reasoning Models (NeurIPS 2026 · similarity 0.87)
• ARES: Multimodal Adaptive Reasoning via Difficulty-Aware Token-Level Entropy Shaping (ICLR 2026 · similarity 0.87)
• Self-correcting for Debiasing Large Language Models (ICML 2026 · similarity 0.86)