Verifiable Rewards Improve Native Visual Reasoning Training
A 300-task procedural suite replaces preference-only judging with task-specific scorers, raising matched reinforcement-learning performance from 0.509 to 0.548.
Sep 4, 20264 min2608.26105
Machine Learning
cs.LGAll aspects of machine learning research including supervised, unsupervised, and reinforcement learning.
Sort:
Million-Clip Dataset Tests Video Reasoning Beyond Visual Quality
VBVR pairs 200 curated reasoning tasks with rule-based, human-aligned evaluation to study whether video models generalize across spatiotemporal reasoning problems.
Aug 30, 20264 min2602.20159
SkillNet Cuts Agent Steps Through Reusable Skills
An ontology, evaluation scheme, and 600,000-skill repository let agents retrieve and compose prior procedures, reporting 40% higher rewards with 30% fewer steps.
Aug 23, 20264 min2603.04448
Agent Swarm Cuts Complex Task Latency Up To 4.5×
Kimi K2.5 couples joint text-vision training with dynamically scheduled parallel subagents, improving agentic search scores while reducing time to target quality.
Aug 13, 20264 min2602.02276
Kimi K3 Brings Open Weights Closer to Frontier Agents
A 2.8T-parameter MoE combines Delta Attention, 16-of-896 expert routing, and agentic reinforcement learning to approach leading proprietary systems at lower task cost.
Aug 10, 20264 min2607.24653
Qwen-CUA Reaches 86.2 on Verified Computer Use
A 397B-A17B mixture-of-experts agent learns screenshot-only keyboard and mouse control from verifiable interactive trajectories, improving OSWorld-Verified performance to 86.2.
Aug 8, 20264 min2608.02352
Request-Level Sharing Cuts Recommendation Inference Cost
ROCS separates request-side and candidate-side computation, reaching up to 3× higher retrieval QPS and 50% higher ranking QPS in production.
Aug 1, 20265 min2607.27744
Mini Activations Narrow Frontier Model Gaps
A sparse 229.9B-parameter MoE activates 9.8B parameters per token and pairs agent-native data with RL for coding, cowork, and reasoning tasks.
Jul 31, 20265 min2605.26494
Hybrid Mamba MoE Trades Dense Scale for Throughput
Soofi S activates 3B of 30B parameters per token and matches larger open models while reaching 4.8k TPS/GPU at 40K context.
Jul 29, 20265 min2607.09424
ABot-World-0 Runs Interactive Worlds on One Desktop GPU
Progressive causal distillation and a co-designed streaming stack produce 720P rollouts at up to 16 FPS on an RTX 5090.
Jul 28, 20265 min2607.19191