norgitov/ trends
Technology · people · ideas
RSS
RUEN
Back to discovery/Daily Papers2 hours ago
ResearchDaily Papers

1% of Tokens Can Be Enough: On Gradient Estimation in On-Policy Distillation

The paper explores sparse on-policy distillation (OPD) by focusing on gradient estimation challenges when allocating teacher supervision to a small subset of tokens. It introduces an information-efficiency ratio (IER) to assess gradient estimation reliability and combines it with existing usefulness scores for better token selection, achieving improved performance with minimal token budgets.

Open original
SIGNAL FROM THE SOURCE
2
source votes
Tracking sinceSeptember 22, 20262 source votes
Momentum—More observations needed
Discussion—Read comments ↗
PublishedSeptember 21, 2026Huanxin Sheng, Zhiling Ye, Haonan Wang
BEHIND THE NUMBERS

How interest changes

Daily Papers

History starts here

The chart will appear after repeat observations. The current metric comes from the source.

2 source votes

Real observations only. History before source connection is not reconstructed.

WHY IT MAY MATTER

This research provides a method to efficiently allocate teacher supervision in on-policy distillation by considering both the usefulness of tokens and the reliability of gradient estimation, which can be useful in scenarios with limited computational resources or data constraints.

A useful discovery?
KEEP EXPLORING

Connected ideas

Explore topic
Daily Papers

Can MiniMax-H3 Reason About the Physical World? An Evaluation of Omni-Modal Generative Model

This paper evaluates the ability of the MiniMax-H3 multimodal generative model to reason about the physical world. The study introduces a new framework for assessing physical reasoning through four dimensions, using tasks that require integrating information across multiple modalities. The model achieves an overall success rate of 41.97% across 517 evaluation instances, with video-based reasoning performing best and audio-based reasoning least effectively.

6 days agoDaily Papers

IntBMoE: Integrating Block-Level Conditioning into Expert Composition for Full-Participation Mixture-of-Experts

IntBMoE introduces a block-conditioned Mixture-of-Experts (MoE) framework that decouples participation, execution, and materialization by combining dense expert composition with sparse block execution. It uses a learned codebook for blocks and a hypernetwork to merge experts, achieving full participation while keeping execution and memory costs low. Experiments show improvements in image classification, language modeling, and sequential recommendation, with real-world deployment in AMap's recommendation system.

4 days agoDaily Papers

Self-Evolving Search Index

The paper introduces SELF-INDEX, a framework that enables an index to self-evolve without human intervention by autonomously diagnosing retrieval issues, revising index keys, and validating changes. It also uses a Query Simulator to proactively explore new retrieval demands, improving performance across various corpora and downstream applications.

5 days agoDaily Papers

ScienceIDE: Turning the World's Scientific Codebase into Agent Learnable Environments

The 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.

6 days agoDaily Papers

Rethinking Critic Learning in PPO: Understanding and Mitigating Value Flattening

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.

6 days agoDaily Papers

ActionPiece: Rethinking Action Tokenization for Autoregressive Vision-Language-Action Models

The paper introduces ActionPiece, a method for autoregressive vision-language-action models that improves action tokenization by preserving physical action relationships through joint supervision of representation learning and quantization. It uses physical rank consistency (PRC) to measure how well tokenization maintains local physical distance rankings after reconstruction, and shows improvements in policy success on multiple benchmarks.

6 days ago