Unverified50% confidenceOpinionExact time
The developer speculates that failure replay stores only the positions of failed games, so nearly all the crowded boards the agent practiced on ended in collisions, causing the value estimate to be pessimistic.
1
Sources
50%
Confidence
Long-term
Relevance
10/29/2026
First Seen
Sources
Related Entities
Related Claims
Unverified级联失败是指单次工具调用错误会污染后续上下文，导致模型在错误的中间状态上继续推理，最终使整个任务偏轨79% similarUnverifiedIn long-horizon agent scenarios, when trajectories are very long and rewards extremely sparse, almost all samples the critic sees early on are failure trajectories, so value estimates become highly homogeneous or even collapse, causing distorted advantage estimates.78% similarUnverifiedWithout re-planning capability, an agent easily retries repeatedly at the same point of failure, forming an ineffective loop.77% similarUnverifiedA reasonable prediction loss combined with 0% downstream control accuracy is highly consistent with representation collapse, meaning the encoder has fooled the self-supervised objective without learning any useful dynamic information.75% similarVerified仅用成功轨迹训练的模型往往存在分布外泛化的严重缺陷，遇到工具返回错误时便丧失纠偏能力75% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1112529API
curl https://kongchang.com/api/v1/knowledge/claims/1112529MCP
get_claim(id=1112529)