The paper introduces ActObs, a method that supervises observation tokens in addition to action tokens during supervised fine-tuning for reinforcement learning. While standard approaches only focus on action tokens, ActObs also predicts observations, leading to better performance in subsequent reinforcement learning tasks. On Qwen3-4B and Qwen3-8B models, ActObs achieves higher pass@k scores and solves more tasks compared to action-only training, with benefits extending to cross-domain code editing. The method maintains more entropy during reinforcement learning and requires less policy movement, keeping the final policy closer to its initial state.
The chart will appear after repeat observations. The current metric comes from the source.
1 source votesReal observations only. History before source connection is not reconstructed.
This method improves reinforcement learning by jointly supervising actions and observations, leading to better task performance and more efficient policy exploration.
The paper introduces XConf, a method for estimating confidence in language models by leveraging the model's accumulated experience. XConf uses past episodes, including tasks, reflections, confidence levels, outcomes, and lessons learned, to inform current confidence estimates through a recall and reflect process. It outperforms existing methods in discrimination and calibration with lower computational cost.
3 days agoDaily PapersLimiX-2 is a new model in the LimiX family that uses the Contextual Mechanism Networks (CMNs) paradigm, pretrained with Context-Conditional Masked Modeling (CCMM). It focuses on learning joint representations of data generation processes rather than target-centric prediction, showing superior performance on tabular data benchmarks and promoting causal awareness.
3 days agoDaily PapersStepAudio 3 Realtime is an audio-language foundation model designed for real-time spoken interaction, featuring a continuous listen-converse-think-act loop. It includes Deep Perception for interpreting user intent, Seamless Duplex for handling audio stream synchronization, and Think-While-Speaking to balance reasoning and latency. The model achieves high performance on benchmarks like MMSU and Artificial Analysis Full-Duplex, along with a task-success rate on τ-Voice.
6 days agoDaily PapersPhysBrain 1.5 is a unified model for understanding physical environments, generating actions, and predicting future states. It uses a vision-language foundation, encodes language responses and motion as sequences, and is pre-trained on human interaction videos. The model achieves strong performance on embodied understanding benchmarks.
4 days agoDaily PapersLynnReal-Omni is a native multimodal video generation framework that combines text-to-video, image-conditioned generation, and structural control into a single model. It includes a 32B shared multimodal diffusion transformer and a 27B Flash version for real-time rendering, with a data pipeline and evaluation method designed for video generation tasks.
4 days agoDaily PapersThe paper introduces ScienceIDE, a framework that transforms scientific code repositories into programmable environments for training scientific agents. These environments enable task generation, execution, and verification, supporting supervised fine-tuning and reinforcement learning. The approach demonstrates improvements in scientific code repair and general-purpose benchmarks.
2 days ago