GVPO++ is a novel post-training method for large language models (LLMs) that integrates the analytical solution of KL-constrained reward maximization into its gradient weighting scheme, offering a unique optimal solution and flexible sampling distributions without relying on importance sampling. This method also extends to on-policy distillation and enables the optimization of a broad family of objectives.

Researchers propose Test-Time Self-Evolving via Reflection (TTSR), a self-evolving framework for adapting large language models during inference using unlabeled test inputs. TTSR improves upon existing methods by addressing bottlenecks in learnable samples and efficient exploration. The framework alternates between a Student and Teacher role, with the Teacher analyzing failed trajectories and synthesizing targeted questions. Experiments show consistent test-time improvements and strong cross-backbone generalization.

Researchers proposed a method to improve high-resolution range profile classification by conditioning on aspect-angle awareness, resulting in an average accuracy gain of 7% and up to 10% improvement.

A new method, ActObs, is introduced to supervise both agent action tokens and observation tokens during reinforcement learning. This approach improves performance in tasks such as code editing and retains more entropy in the policy, leaving it closer to its initialization.

Researchers propose a method called LLM-as-an-Improver, which uses verification feedback to generate and reselect improved candidates for large language models. This approach, called Verify--Repair--Reselect (VRR), can improve LLM performance and recover correct solutions even when the initial pool is incorrect.

Researchers propose a framework to improve compositional reasoning in language models using reinforcement learning post-training. They find that training on composed tasks leads to better transfer to decomposed tasks, and provide theoretical explanations and evaluations under various conditions.

The paper presents Flare, a lightweight diagnostic model for predicting line-level suspiciousness signals to help large language models refine generated code. Experiments show that Flare outperforms baseline methods, achieving an average improvement of 8.50% with candidate search.

Researchers introduce ActionPiece, a method for preserving physical action relationships in autoregressive vision-language-action models through joint supervision of representation learning and quantization. This approach improves action tokenization and policy learning, achieving high accuracy on various benchmarks.

Researchers propose Mixture of Reused Experts (MoRE), a hybrid model architecture that combines the benefits of Mixture of Experts (MoE) and weight-sharing. MoRE shares expert pools across adjacent layers, improving efficiency and performance while maintaining competitive language modeling capabilities.

Researchers explored procedural pretraining for molecular property prediction, finding it can improve performance even after molecular pretraining. A three-stage training pipeline was used, consisting of procedural pretraining, molecular pretraining, and downstream fine-tuning. Results showed a 4.8% reduction in test error for Lipophilicity, and the benefit was strongest under data scarcity.

Researchers propose a new approach for Text-to-SQL pipelines that uses large language models to eliminate the need for schema linking, improving accuracy and efficiency. This development is relevant to people building or operating AI agents as it impacts the design and optimization of Text-to-SQL systems.

A new learning framework, PIVOT, is proposed to improve the reasoning capabilities of large vision-language models by anchoring policy optimization around informative visual reasoning signals. This is achieved through a self-calibrated experience replay mechanism and a vision-guided advantage allocation mechanism.

Researchers introduce Attention Calibration for Position-Fair Dense Retrieval, a method to counteract positional skew in dense retrieval. This method improves retrieval effectiveness and reduces position sensitivity by 43% in 16 evaluated combinations. The authors release their code on GitHub.

A new method, FADE, has been proposed to mitigate hallucinations in Large Vision-Language Models (LVLMs) by reducing language-prior dominance. This is achieved through a training-free approach that attenuates the outputs of feed-forward networks (FFNs), which are found to be the source of language priors. Evaluations show that FADE effectively reduces hallucinations while preserving inference efficiency.

A research paper introduces ProtoLIP, a lightweight prototype-mediated evidence layer for query-conditioned vision-language models. It improves evidence localization and separation without requiring spatial annotations or backbone retraining, and remains competitive with a spatially supervised grounding model.

Researchers propose Turn-level Multiscale Density Ratio Estimation (tlm-DRE), a training method for Large Language Model (LLM) agents that improves performance on complex multi-turn tasks. The method assigns different weights to turns and trains on token-level gaps across multiple turns. Experiments show competitive results compared to traditional alignment methods.