Alibaba's Accio team, with Fudan University and Shanghai Innovation Institute, introduces Neos, a unified training framework that couples data construction, process rewards and environment co-evolution. Its 9B and 35B models approach frontier closed-source performance on four multimodal search benchmarks by building multi-hop tasks with visual checkpoints and filtering out shortcut solutions.
Credit Assignment
信用分配
Problem of determining which step is responsible for the final outcome in multi-step decisions
Curated Intelligence
Researchers from Tsinghua University and Tencent propose TreeOPSD, which improves credit assignment in long-horizon agent reinforcement learning via tree-based trajectory sampling and turn-level credit estimation. On Qwen3-30B-A3B-Thinking-2507, it achieves 21.20% average Exact Match across four Deep Research benchmarks, a 4.97-point gain over GRPO's 16.23%.
To address error propagation across multi-stage reasoning in empathetic spoken dialogue, researchers from CUHK and Microsoft introduce EchoChat, a unified framework that organizes perception, theory-of-mind inference, strategy planning and speech response generation, constrains inter-stage dependencies with cascaded gated rewards, and applies step-level credit assignment for targeted updates. Accepted at NeurIPS 2026, the work reframes empathy as a decomposable cognitive process rather than a direct input-to-response mapping.
Amazon releases a hands-on guide for Multi-Turn Reinforcement Learning (MTRL) on SageMaker AI to address the limitations of supervised fine-tuning in agent tasks. The solution trains models using sparse reward signals from final task outcomes by connecting to Bedrock AgentCore or custom infrastructure, enabling effective learning for complex multi-step interactions.
Google researchers propose a new automated framework to address semantic drift and cascading failures in agentic long-form video generation pipelines. By optimizing prompt structures and error tracing mechanisms, the approach significantly improves cross-shot character consistency and narrative coherence while reducing manual intervention.