Understanding a language model's decision requires knowing how its internal computations affect its preference between answers. Gradient-based methods estimate this influence by predicting the effect of changing those computations. In mixture-of-experts (MoE) models, replacing one expert can change which experts later layers select. Those switches can alter the answer scores in ways that derivatives do not capture. We introduce a predictor that uses the model's own expert outputs and selection rules to estimate these extra effects, then adds them to a second-order prediction. Across questions from three benchmarks, our method reduces the error in predicting answer-score changes by 60\% on Qwen3.6 and 47\% on Gemma4, relative to a multistep baseline using gradients and second derivatives.
Zhen Xu, Hang Yin, Zitao Zhang, Kevin Zhao, Xuefeng Liu
原文摘要
Agents are increasingly built to learn from experience: they keep conversation histories, memory files, retrieved notes and skills, and reuse what worked before. We show that this experience turns against them when their tools change silently. We test a range of models with tools defined as in the Model Context Protocol (MCP). After a silent change, experienced agents fail on most tasks, while agents new to the task mostly succeed. We call this failure Experience Rot. Experienced agents copy their earlier call instead of checking the current interface, even when the tool definition shows the change. Then they report success. When the documentation is incomplete, experience is what makes agents succeed. After a change, the same experience leaves them worse off than no experience at all. The failure appears in every form of memory we test and grows with the amount of experience. Simple remedies fall short: announcing the update barely helps, and pointing the agent to the documentation fails whenever the documentation does not spell out the answer.
Real-world robot deployment requires adapting learned policies to unseen obstacles without collecting new demonstrations or retraining. Existing policy-level guidance often applies obstacle corrections independently at individual horizon steps, which may address local obstacle violations but fail to coordinate the action sequence into a complete, task-directed detour. How obstacle responses should be propagated across the action horizon to coordinate complete, task-directed avoidance motion remains underexplored. We introduce an inferencetime obstacle-guidance framework for pretrained relative-action diffusion policies. It rolls out cumulative relative actions into a workspace trajectory, evaluates obstacle interactions along the trajectory with an SDF-based obstacle model, and propagates each obstacle response to all contributing action increments during denoising. Real-robot experiments across seven obstacleconstrained scenarios spanning four manipulation tasks show that our method achieves 105/140 (75.0%) collision-free task successes, compared with 8/140 (5.7%) for absolute-action pointwise guidance. Ablations show that the relative-action horizoncoupled formulation provides the larger observed improvement across both obstacle models.