2026· Annual Meeting of the Association for Computational Linguistics· pp. 22104-22121· 1 citation· 44 references
Computer Science
TL;DR
To mitigate a critical imbalance during the exploration-and-learning process, this work approaches head-tail re-balance during the exploration-and-learning process from two perspectives: distribution-reshaping and trajectory-resampling.
Abstract
Self-improvement has emerged as a main-stream paradigm for advancing the reasoning capabilities of large vision–language models (LVLMs), where models explore and learn from successful trajectories iteratively. However, we identify a critical imbalance during this process: the model readily generates high-quality trajectories for simple queries (i.e., head data) but struggles with complex ones (i.e., tail data). This bias drives the optimization to disproportionately prioritize simple reasoning skills, while inhibiting the acquisition of complex capabilities. As iterations progress, this imbalance becomes more acute—a dynamic we term the “Matthew effect” 1 , ultimately stalling performance gains. To mitigate this, we approach head-tail re-balance during the exploration-and-learning process from two perspectives: distribution-reshaping and trajectory-resampling. Extensive experiments on Qwen2-VL-7B-Instruct and InternVL2.5-4B models across visual reasoning tasks demonstrate that our methods consistently improve visual reasoning capabilities, outperforming vanilla self-improvement baselines by an average of 3.86 points. 2
Lookahead Optimization for Rehearsal (LOR) is introduced, establishing a new and more robust paradigm for rehearsal-based OCIL and significantly out-performs state-of-the-art methods.
The pursuit of autonomously self-improving models has attracted growing interest in the era of large-scale foundation models. Drawing inspiration from the concept of"enlightenment"or"aha moment"in human brain, we hypothesize that large models exhibit an analogous enlightenment phenomenon-a latent capacity for sudden capability boost. Then, we propose Enlightenment, a novel training-free post-tuning paradigm for large-scale models. Our approach modifies shortcuts for key modules/layers without weight updates, while existing training-free ones predominantly manipulate attention weights. We introduce two architecture-specific instantiations: i) For large language models, we propose attention head-mixing shortcuts that recalibrate attention weights by linking the initial attention head's output to all other target heads, modulated by an adaptive scaling factor initialization strategy. ii) For vision-language models, we apply a lightweight scalar-modulated factor to residual connections in the decoder layers, regulating information flow. Extensive experiments show that Enlightenment efficiently unlocks the latent potential of pre-trained networks, yielding remarkable performance improvements across diverse benchmarks and models.
Jingxiao Liao, Tianwei Zhang, Yu-Hao Jiang et al.· 0 citations
A systematic literature review on how RL are adapted and scaled as a fundamental post-training tools and how innovations in the RL pipeline enhance the domain-specific LLMs is conducted.
Qianyue Hao, Lin Chen, Xiaoqian Qi et al.· ACM Computing Surveys· 1 citation
We present MAJEPPA, a self-supervised framework to learn piano performance representations that span the full skill spectrum, from beginner practice sessions to virtuoso concert recordings. We curate the MAJEPPA dataset, comprising ~4,000 annotated recordings across six expertise levels and six recording contexts. We adapt a single pre-trained MIDI autoregressive model with a joint objective: next-token prediction learns score-conditioned performance generation at various skill levels, while InfoNCE and supervised contrastive losses align abstract score and performance representations in a joint embedding space. The proposed model both generates and understands performances in a unified framework. By introducing the EVPMR benchmark, a suite of downstream tasks spanning quality assessment, competition ranking, mistake and technique classification, we evaluate the learnt representations, demonstrating progress towards a real-world model for the piano performance space.
Jinwen Zhou, Huan Zhang, Weixin Zhai et al.· 0 citations
Vision-Language-Action (VLA) models acquire broad embodied capabilities through large-scale pretraining, yet their generalization remains far more fragile than that of LLMs and VLMs. The prevailing remedy, post-training via supervised fine-tuning or reinforcement learning, improves task-specific performance but narrows the generalist capability that makes pretraining valuable. We identify a key bottleneck: VLA failures stem not only from action generation but also from action evaluation. A diagnostic pass@k study confirms that frozen VLAs already contain competent behaviors in their output distribution, with overall success rates rising from 33% at pass@1 to 92% at pass@32. Inspired by this, we propose SVA (Search, Value, and Act), a simple framework that equips frozen VLA policies with long-term consequence awareness. SVA first uses Monte-Carlo tree search in simulation to fully explore the VLA's output distribution and collect diverse trajectories annotated with empirical returns; this knowledge is then distilled into a lightweight Q-value model that predicts the expected consequence of candidate actions; at deployment, the frozen VLA proposes multiple candidates and the evaluator selects the one with the highest uncertainty-regularized Q-value, requiring no simulator access. By decoupling action proposal from consequence evaluation, SVA preserves the generalization capacity of the VLA backbone while substantially improving task success rates. Experiments across embodied benchmarks show that SVA consistently improves generalization on unseen tasks and exhibits strong test-time scaling behavior. Strikingly, SVA enables a 9B VLA to outperform a 27B VLA by 7 points at 27% lower inference latency, suggesting that scaling test-time evaluation is more cost-effective than scaling model size.
Xinyi Xie, Zican Hu, Zhanyun Liu et al.· 0 citations
Multimodal Large Reasoning Models (MLRMs) have achieved strong performance on tasks requiring visual understanding and multi-step inference. However, as reasoning trajectories grow, models may become less effective at using information established earlier in the context, increasing the risk of reasoning errors. Existing approaches primarily address this problem by sustaining visual grounding throughout reasoning. However, reasoning also transforms visual observations into task-specific relations, constraints, and intermediate conclusions whose influence may weaken over long trajectories. Our attribution analysis suggests that correctness is not consistently separated by image attribution alone, but is more closely associated with whether trajectories retain and integrate such reasoning-derived information across stages. Motivated by this, we introduce TRAM (TRajectory-derived Auxiliary Memory), a training-free method that augments standard decoding with an auxiliary memory pathway derived from the model's own reasoning trajectory. TRAM consolidates completed reasoning into a compact latent memory, updates it online through fast and slow recurrent streams, and feeds it back into selected decoder layers through a lightweight residual pathway. Experiments across four MLRM variants on eight benchmarks show that TRAM improves performance over vanilla decoding on mathematical, scientific, and general visual reasoning tasks without additional training.
Kang Liu, Zijing Wang, Yongkang Liu et al.· 0 citations