The paper identifies a problem called Value Flattening in Proximal Policy Optimization (PPO) critics, where estimated state values change sharply while critic predictions remain flat. It introduces SP^3O, a method that applies value loss to only a few states per response, showing improvements in policy learning.
The chart will appear after repeat observations. The current metric comes from the source.
33 source votesReal observations only. History before source connection is not reconstructed.
This research addresses a critical issue in PPO that affects the accuracy of value estimation, which is essential for effective reinforcement learning in large language models.
StepAudio 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.
5 days agoDaily PapersStepAudio 3 Music is a large-scale music generation model that uses explicit musical planning and text control. It employs a flow-matching diffusion Transformer to predict audio latents, with a discrete-continuous design and a Mixture-of-Experts autoregressive model for structured music generation.
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.
3 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.
3 days agoDaily PapersThe 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.
2 days agoDaily PapersZGCM-1 is a 7B parameter foundation model designed for math and agentic search, emphasizing efficiency through architectural and system co-design, progressive curriculum learning, and an AI-native R&D workflow. It demonstrates competitive performance against larger models on mathematical and search tasks, with improvements in training efficiency and open-source availability of training materials.
6 days ago