The paper revisits large‑language‑model (LLM) based automatic speech recognition (ASR) from an entropy‑allocation perspective, introducing three metrics to analyze how entropy reduction is distributed between the speech encoder and the LLM. It proposes a multi‑stage training strategy that addresses entropy‑allocation inefficiencies, including a redesigned pretraining phase to reduce the speech‑text modality gap and an iterative asynchronous fine‑tuning stage to maintain functional decoupling and limit encoder drift. Experiments on Mandarin and English benchmarks demonstrate competitive performance with state‑of‑the‑art models using only 2.3 B parameters, while also reducing hallucinations through the decoupling‑oriented design.
By Yuan Xie, Jiaqi Song, Guang Qiu, Xianliang Wang, Ming Lei, Jie Gao, Jie Wu
The paper investigates whether language models can be fine‑tuned to produce outputs that better match intended probability distributions, even when the valid outputs cannot be enumerated in advance. Two calibration fine‑tuning methods—soft‑target and hard‑target—are evaluated across 12 models from four families, showing substantial improvements in structured‑sampling fidelity on unseen distribution families and parameter settings. The gains also transfer to broader stochastic generation tasks, expanding first‑token support and improving NoveltyBench scores, though some downstream capabilities such as arithmetic reasoning may be slightly affected.
By Davide Baldelli, Sruthi Kuriakose, Maryam Hashemzadeh, Amal Zouaq, Sarath Chandar
EmoGavel is a reinforcement learning framework that incorporates verifiable rewards to improve multimodal emotion preference learning. It introduces grounded supervised fine‑tuning (Grounded SFT) to jointly learn preference prediction, error localization, repair, and missing‑evidence recovery, and then applies grounding‑aware group relative policy optimization (GA‑GRPO) to prioritize preference correctness and recover learning signals from all‑wrong groups. Experiments on EmoPrefer and controlled‑error evaluation sets demonstrate that EmoGavel enhances natural preference prediction, controlled‑error discrimination, and maintains strong error‑grounding performance.
By Zilong Huang, Kong Aik Lee, Zezhong Jin, Man-Wai Mak, Tatsuya Kawahara
arXiv:2610. 13653v1 Announce Type: cross Abstract: Reinforcement learning with verifiable rewards (RLVR) has substantially improved LLM reasoning, yet growing evidence suggests it mainly performs distribution sharpening: concentrating probability mass on reasoning paths the base model already produces.
By Abiel Malkani Talwar, Christopher Nemeth, Henry Moss
The paper introduces an uncertainty-guided progressive contraction framework to reduce the amount of reasoning supervision needed for large language models. By building Monte Carlo reasoning trees around intermediate states, the method contracts the training corpus into a compact teacher-informed pool and then performs learner-aware contraction during supervised fine-tuning. Experiments show that the approach can use as little as 0.22% of the optimization budget of full-data SFT while maintaining or improving reasoning performance across diverse benchmarks and model families.
By Suorong Yang, Jiaxin Wang, Jiaming Su, Chentao Yue, Furao Shen, Peng Ye, Soujanya Poria
The paper introduces TQS, a Trajectory Quality Score that evaluates coding-agent interaction trajectories based on structural signals such as information focus, intent–action alignment, and policy dynamics, rather than just final outcomes. TQS operates in linear time without requiring an LLM judge and is applied to 155,144 successful trajectories from a public corpus. Fine‑tuning agents on trajectories ranked highest by TQS consistently outperforms models trained on random or low‑ranked subsets on the SWE‑Bench‑Verified benchmark, demonstrating the value of process‑level quality filtering.
By Sida Li, Yuanxin Liu, Shicheng Li, Linli Yao, Xu Sun, Haoli Bai, Xiaohui Li
The paper introduces ChemDiffAgent, a masked diffusion language agent designed for tool‑integrated chemical reasoning. Unlike traditional autoregressive agents, it treats reasoning as iterative denoising over interleaved reasoning and tool‑use trajectories, allowing joint decision making under full‑trajectory context. A two‑stage post‑training framework further refines tool use and reasoning, and the authors benchmark the model on six chemistry tasks, reporting a ~20% performance gain over autoregressive baselines and competitive results against larger models.
By Mengdi Liu, Chenghao Jia, Hong Chang, Shiguang Shan
Neurons++ is a visual‑stream framework that reconstructs both static images and dynamic videos from fMRI data. It builds on the earlier Neurons model by extending from video‑only to image‑and‑video reconstruction and replacing parallel task conditioning with a serial, cortex‑inspired decoding pipeline. The approach first pre‑trains on image‑fMRI pairs for stable spatial‑semantic initialization, then fine‑tunes separately for image and video tasks, organizing fMRI representations into spatial, semantic, language‑level, and temporal cues, and achieves state‑of‑the‑art video reconstruction with competitive image results while enabling neural‑to‑visual interpretability and ROI‑guided semantic modulation.
By Haonan Wang, Yuchen Wang, Jingyu Lu, Qixiang Zhang, Lehan Wang, Xuanqi Huang, Xiaomeng Li
Mesh‑Native Tokenisation introduces Mesh‑Native Convolution (MNC), a locality‑aware tokenisation layer that encodes unstructured mesh‑based data similarly to convolutional neural networks. This allows physics foundation models trained on grid data to be retrofitted to arbitrary unstructured meshes with minimal retraining, achieving competitive accuracy even in zero‑shot deployment. The paper provides theoretical validation of MNC’s consistency with classical CNN encoders and demonstrates its superior performance over existing baselines after only a few fine‑tuning epochs.
By Georg Maierhofer, Miles Cranmer, Payel Mukhopadhyay
The paper introduces w$^{2}$VLA, a Vision‑Language‑Action model that restructures information flow to separate declarative knowledge (concepts and entity semantics) from procedural knowledge (how to act). By modulating the robot state sequence with visual, spatial, and skill information in a compositional, interpretable way, the model achieves robust behavior cloning and unprecedented zero‑shot skill transfer to unseen objects, overcoming brittleness seen in current state‑of‑the‑art VLA models.
By Nikolaos Tsagkas, Andreas Sochopoulos, Chris Xiaoxuan Lu, Oisin Mac Aodha, Alexandros Kouris
AgroGround is a large‑scale dataset that links plant disease and agricultural target labels to image regions, enabling models to both recognize and localize these targets. It was created by automatically converting eight agricultural VQA datasets into 794,850 grounded instruction examples, including healthy images for negative supervision. Fine‑tuning a shared vision‑language model on these instructions improves recognition, joint accuracy, and healthy‑image abstention, outperforming its annotation teacher on both the AgroGround benchmark and the external PlantSeg test set.
By Abdulla Alshehhi, Zongyan Han, Rao Muhammad Anwer
The paper demonstrates that hidden capabilities in AI models can be provably concealed by embedding encrypted neural networks as backdoors within the host model’s weights. These hidden circuits are designed to decrypt and activate only when a secret trigger is provided, and they remain difficult to elicit or interpret even with full access to the model weights, thanks to cryptographic guarantees. The authors implement these constructions in PyTorch and test their resistance against various elicitation techniques on a chemical reaction prediction task, showing that current audit methods cannot reliably uncover such hidden behaviors.
By Andis Draguns, Sumeet Ramesh Motwani, Raymond Douglas, Christian Schroeder de Witt
SWEAR is a difficulty‑aware reasoning framework for large audio‑language models that derives five‑dimensional difficulty scores and compressed reasoning traces from offline assessments. It learns these difficulty signals during supervised fine‑tuning and employs reinforcement learning to adapt reasoning allocation across stages, shortening reasoning for easy instances while maintaining performance on hard ones. Experiments on MMAU‑test‑mini and MMAR demonstrate that SWEAR improves accuracy and reduces generated tokens.
MotionCoach is a multimodal agent designed to enhance video editing models for professional motion editing, such as sports and martial arts. It generates grounded prompts and reference clips through multi‑turn tool use and self‑verification, enabling precise motion guidance without retraining the base editing models. The system is trained on a curated dataset and evaluated on VPMEBench, achieving significant score improvements on benchmark datasets.
The paper investigates whether the adaptive circuit‑growth method Q‑FLAIR, which stops adding gates when training loss ceases to improve, can be combined with Caro et al.’s bound on generalization to predict how circuit depth should scale with training data size. Experiments on full‑resolution MNIST 3‑vs‑5 classification with training sets from 2,000 to 10,000 samples show no clear relationship between dataset size and the final circuit size; both circuit size and test accuracy vary non‑monotonically, and seed‑to‑seed variance is large. While the empirical generalization gap never exceeds Caro et al.’s bound in 14 of 15 runs, the gap correlates only weakly with the bound’s value, indicating that the bound’s predictive power is limited.
By Saeefa Rubaiyet Nowmi, Md Mahmuduzzaman Kamol, Mohammad Saidur Rahman
AgentFly is an agentic reinforcement learning framework that introduces a unified resource layer to manage heterogeneous environments—such as sandboxes, model services, and external APIs—through a single scheduling engine. The framework is organized into four layers: an agent layer that abstracts agents, tools, and rewards; a rollout layer that composes agent loops and computes rewards; a context layer that organizes rollouts and injects contextual information; and a low‑level resource layer that handles resource management. AgentFly offers prebuilt tools and environments, demonstrates successful agent training across multiple tasks and models, and provides the first controlled cross‑framework throughput comparison against other agentic RL frameworks.
By Renxi Wang, Rifo Ahmad Genadi, Bilal El Bouardi, Yongxin Wang, Fajri Koto, Zhengzhong Liu, Timothy Baldwin, Haonan Li
SatFix is a feed‑forward framework that performs absolute visual localization of UAVs in satellite maps using a single oblique image or a short multi‑view clip. It replaces traditional cross‑view retrieval by regressing a continuous 3‑DoF pose—2D position and heading—directly from satellite‑grid features without requiring a 3D map or auxiliary sensors. On the newly introduced University‑Metric dataset, SatFix localizes over half of test frames within 50 m and achieves median errors of 45.66 m and 20.81° with a single view, improving further with multiple views.
By Jiarui Zeng, Kun Shi, Chiman Vong, Zhedong Zheng
Project Greenhouse explores building fully open and sovereign agentic search models with modest computational resources. The authors present a competitive pointwise decoder-only reranker trained from scratch and fine-tuned on common datasets, avoiding reliance on third‑party open‑weight backbones. They demonstrate that most experiments can be run on a handful of GPUs and release data, code, configurations, and checkpoints for reproducibility.
By Jimmy Lin, Sahel Sharifymoghaddam, Lingwei Gu, Nour Jedidi
NanoProof is the first factorized execution‑guided theorem prover in Lean 4 whose training data, tooling, pipeline, and weights are all publicly released, enabling end‑to‑end reproducibility with open‑source resources. The authors provide a dataset of structured proof trees and a programmatic interaction tool for Lean 4, emphasizing compute efficiency to make training and evaluation more accessible. NanoProof achieves a 50.8% pass@16 on MiniF2F‑Test, outperforming comparable systems while using far less compute, demonstrating that such provers can be rebuilt from scratch with modest resources.
By Mat\v{e}j Kripner, Milan Straka
The study investigates whether large language models (LLMs) rely solely on their initial task context or adapt based on experimental feedback when selecting fine‑tuning configurations for neural operators. Experiments across partial differential equation families show that LLMs outperform random search and Bayesian optimisation, with early configurations already near the top of a random‑search pool and subsequent proposals changing in response to validation scores. Controlled interventions confirm that the LLM combines a task‑dependent prior with sensitivity to observed outcomes.
By Julian Chan, Javier Mora Jimenez