Multimodal Continuous Reasoning via Asymmetric Mutual Variational Learning
Summary
Proposes Asymmetric Mutual Variational Learning (AMVL) to resolve train-inference mismatch in multimodal continuous reasoning by using bidirectional calibration to prevent answer leakage and improve latent-space stability, achieving significant gains on the BLINK benchmark.
View Cached Full Text
Cached at: 07/02/26, 07:47 AM
Paper page - Multimodal Continuous Reasoning via Asymmetric Mutual Variational Learning
Source: https://huggingface.co/papers/2607.00461
Abstract
Asymmetric Mutual Variational Learning addresses train-inference mismatch in multimodal reasoning by using bidirectional calibration to prevent answer leakage and improve latent-space stability.
Multimodal Large Language Models(MLLMs) are often constrained by alanguage-space bottleneck, forcing complex visual reasoning into discrete tokens which can lose perceptual nuance. A promising alternative iscontinuous latent reasoning, where the goal is to discover implicit reasoning pathways that bridge the multimodal query and the final answer. However, this introduces a severetrain-inference mismatch: a training-timeposterior, conditioned on the ground-truth answer, can exploit answer-dependent shortcuts. Standardvariational trainingthen forces the inference-timepriorto mimic aposteriorthat has access to information unavailable at test time, leading to poor performance. To address this, we propose Asymmetric Mutual Variational Learning (AMVL), a framework that resolves this mismatch via abidirectional calibrationobjective. Aforward KL divergencetrains the target-agnosticpriorto match theposterior, while a novelreverse KL divergencesimultaneously regularizes theposterior, preventing it from collapsing into inference-incompatible regions and mitigating this ``answer leakage’’. We provide theoretical analysis formalizing this leakage aspriorcontamination and prove that our dual-KL objective reduces it. We instantiate AMVL in alatent-integrated MLLMand show that it consistently outperforms strong discrete and latent-reasoning baselines, improving the average score on the complexBLINK benchmarkby +10.83 and achieving gains of up to +32.00 on individual reasoning tasks, with analyses confirming improved latent-space stability.
View arXiv pageView PDFAdd to collection
Get this paper in your agent:
hf papers read 2607\.00461
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2607.00461 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2607.00461 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2607.00461 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
LEMUR: Latent Entropy-aware Multimodal Unlearning via Visual-anchored Reasoning Redirection
This paper identifies a privacy vulnerability in RL-trained multimodal large reasoning models, which can leak sensitive facts in their reasoning traces even after unlearning, and proposes LEMUR, a training-free inference-time framework that uses entropy dynamics to detect and suppress such leakage.
Eliciting Complex Spatial Reasoning in MLLMs through Wide-Baseline Matching
This paper introduces ReasonMatch-Bench, a benchmark for wide-baseline matching in multimodal LLMs, and proposes Dynamic Correspondence Reinforcement Learning (DCRL) to improve spatial reasoning. Experiments show significant gains on the benchmark while maintaining general performance.
DeepVoyager-VL: Incentivizing Vision-in-the-Loop Search for Long-Horizon Multimodal Agents
DeepVoyager-VL proposes a long-horizon multimodal deep-search framework that integrates visual evidence into intermediate reasoning, using a multimodal event graph for data synthesis and fine-tuning without reinforcement learning, achieving strong performance across ten benchmarks.
VLMs are Good Teachers for Video Reasoning via Adaptive Test-Time Optimization
This paper introduces a paradigm where Vision-Language Models (VLMs) act as test-time teachers to guide Video Generation Models (VGMs) via differentiable rewards and LoRA optimization, achieving a 16.7-point average improvement on video reasoning benchmarks.
SVR-R1: Bootstrapping Multi-modal Reasoning with Self-verification in Reinforcement Learning
Introduces SVR-R1, a multi-turn reinforcement learning framework that uses the model's own verification as a learning signal for multi-modal reasoning, achieving significant accuracy improvements over standard GRPO baselines on vision-language reasoning benchmarks.