5arXiv cs.AI (Artificial Intelligence)·
ReToken: Single learnable embedding improves VLM visual retrieval in long contexts
ReToken introduces a single learnable embedding trained as an explicit retrieval target to select sparse, query-relevant visual tokens from a pre-filled KV cache, addressing performance degradation in vision-language models under long visual contexts. Trained on a small image-QA dataset, it improves Qwen3VL-8B by 13.4 points on Visual Haystacks and transfers zero-shot to long video benchmarks with an 8.0-point gain on LVBench. The method is lightweight enough for both training and long-video inference on a single H100 GPU. Code is publicly released.
Related guides (3)
Related events (8)
Reroute: Training-free recoverable visual token routing for vision-language models
A new arXiv preprint proposes Reroute, a training-free plug-in that replaces the standard rank-and-remove visual token pruning paradigm in VLMs with a recoverable routing mechanism. Instead of permanently discarding low-ranked tokens, Reroute defers them to re-enter the candidate pool at later decoder stages, addressing the problem that token importance shifts across decoder depth. Evaluated on LLaVA-1.5 and Qwen backbones augmented with FastV, PDrop, and Nüwa pruning methods, Reroute improves grounding performance under aggressive token reduction without sacrificing general VQA accuracy. The approach preserves the theoretical compute and KV-cache budget of the underlying pruning method.
VEPO: Vision-anchored token selection improves RL for visual reasoning
A new arXiv paper identifies a failure mode of entropy-based credit assignment in multimodal reinforcement learning: vision-sensitive tokens with naturally low entropy are systematically ignored, causing the mechanism to collapse in visual reasoning tasks. The authors propose VEPO (Vision-Entropy token-selection for Policy Optimization), which couples visual sensitivity with token entropy via a multiplicative scheme to redirect gradient credit toward tokens that are both visually grounded and semantically informative. VEPO outperforms entropy-only baselines by 2.28 points at 7B scale and 3.15 points at 3B scale on visual reasoning benchmarks.
VRRL: Reinforcement learning framework for visually grounded self-reflection in vision-language models
Researchers propose VRRL, a reinforcement learning training framework designed to improve self-reflection in large vision-language models (LVLMs) by ensuring corrections are grounded in visual inputs rather than purely textual reasoning. The framework uses two mechanisms: random masking of trajectory prefixes to emphasize error recovery, and buffered roll-ins from an experience replay buffer to expose models to diverse failure states. Evaluated on visual grounding tasks (tables, charts) and spatial navigation benchmarks, VRRL substantially improves out-of-distribution accuracy over standard RL and reflection-oriented fine-tuning baselines. The work addresses a concrete failure mode where LVLMs lose visual grounding during chain-of-thought reflection.
DepthWeave-KV: Token-adaptive cross-layer KV cache compression for long-context inference
A new arXiv preprint introduces DepthWeave-KV, a KV cache compression method that factorizes key-value states across neighboring transformer layers using shared low-rank channel bases while retaining token-specific residuals for attention-sensitive positions. A token-conditional depth router allocates higher reconstruction rank to instruction-bearing and retrieval-critical tokens, with calibration-free online error tracking during generation. The method achieves 8.3x KV memory reduction at 64K context while maintaining near-full-cache quality on LongBench, Needle-in-a-Haystack, and L-Eval benchmarks. The work addresses a practical bottleneck in long-context inference without requiring base model retraining.
Qwen2.5-VL-32B: Reinforcement-Learning-Optimized Vision-Language Model Released
Alibaba's Qwen team has released Qwen2.5-VL-32B-Instruct, a 32-billion-parameter vision-language model built on the Qwen2.5-VL series and further optimized with reinforcement learning. The model is open-sourced under the Apache 2.0 license and available on Hugging Face and ModelScope. It follows the January 2025 launch of the broader Qwen2.5-VL series, positioning the 32B scale as a balance between capability and deployment practicality.
RELEX: Extrapolating LLM RLVR Training via Rank-1 Parameter Trajectories
This paper demonstrates that RLVR weight update trajectories are extremely low-rank and near-linearly predictable, with a rank-1 approximation capturing most downstream performance gains. The authors propose RELEX, a compute-efficient method that observes a short training window, estimates the rank-1 subspace, and extrapolates future checkpoints via linear regression—requiring no additional training. Evaluated on Qwen2.5-Math-1.5B, Qwen3-4B-Base, and Qwen3-8B-Base, RELEX matches or exceeds full RLVR performance using as few as 15% of training steps, and can extrapolate up to 10–20× beyond the observed prefix. The authors attribute the method's effectiveness to a denoising effect from rank-1 projection that discards stochastic optimization noise.
DelTA: Discriminative Token Credit Assignment for RLVR Training
DelTA introduces a discriminative token credit assignment method for reinforcement learning from verifiable rewards (RLVR) that addresses the problem of high-frequency formatting tokens dominating policy gradient updates. The method estimates per-token coefficients to amplify side-specific gradient directions and downweight shared or weakly discriminative ones, making the effective update direction more contrastive. On seven mathematical benchmarks, DelTA outperforms same-scale baselines by 3.26 and 2.62 average points on Qwen3-8B-Base and Qwen3-14B-Base respectively, with additional gains on code generation tasks.
TEVI: Sparse autoencoders for text-conditioned editing of CLIP image embeddings to improve vision-language alignment
TEVI is a framework that uses sparse autoencoders to disentangle CLIP image embeddings and a learned masking module to selectively reconstruct embeddings conditioned on a given caption, addressing the information imbalance between images and their captions. The approach improves image-text retrieval on both coarse-grained benchmarks (MS COCO, Flickr) and fine-grained long-caption benchmarks (IIW, DOCCI), with larger gains on richer captions. The work also shows improved robustness on the RoCOCO benchmark.