cs.LGNov 18, 2026

Online State-Representation Adaptation under Synthetic Interaction-Cue Remapping

Authors: Bin Yin

Organizations: School of Psychology, Fujian Normal University

Abstract

When cues change what they predict, continuing to update a learned mapping may help an agent choose the appropriate action. We tested this in a one-step synthetic task by comparing a logistic state predictor that continued to learn from state labels with the same predictor frozen at transfer. We evaluated ten conditions in six held-out environments, with 1,296 paired runs per condition and 1,296,000 transfer records in total. Late action accuracy was 93.39% with online updates and 64.63% with frozen coefficients, a paired difference of 28.76 [8.73, 53.19] percentage points (95% environment-bootstrap interval). A direct policy reached 95.65% using the same predictions with a different action threshold. A contextual bandit reached 94.23% and earned more cumulative utility than the adaptive predictor. The bandit used the reward of its chosen action, which in this task reveals the true state when combined with that action. Weights based on development history lowered cumulative utility relative to neutral weights (difference -3.45 [-4.48, -2.43]). The adaptive predictor outperformed its frozen counterpart in late accuracy in each of the six test environments. The other comparisons show how the choice of action threshold and the information available in feedback affect what can be attributed to online adaptation.

Figures & tables

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Apr 6, 2026

Data Attribution in Adaptive Learning

Machine learning models increasingly generate their own training data -- online bandits, reinforcement learning, and post-training pipelines for language models are leading examples. In these adaptive settings, a single training observation both updates the learner and shifts the distribution of future data the learner will collect. Standard attribution methods, designed for static datasets, ignore this feedback. We formalize occurrence-level attribution for finite-horizon adaptive learning via a conditional interventional target, prove that replay-side information cannot recover it in general, and identify a structural class in which the target is identified from logged data.
Sep 18, 2026

Sandwich-Residuals: Parameter-Efficient Test-time Adaptation of World Models

Latent world models enable planning by predicting the effects of actions in a learned representation space, but their predictions can become unreliable when test-time conditions differ from training. Existing test-time adaptation methods address this by updating parts of the pretrained model, often modifying millions of parameters and requiring a choice of which internal components to adapt. We introduce Sandwich-Residuals, a lightweight alternative that keeps the pretrained world model frozen and learns only small residual corrections around the predictor. The residuals are optimized online using the model's self-supervised prediction error and require no rewards, labels, or source-domain data. Across 21 conditions on the AdaJEPA benchmark, our method achieves 1.3×1.3\times the success rate of the frozen model while retaining 95% of the performance of the strongest AdaJEPA variant and adapting 97-99% fewer parameters. Under compound shifts, this advantage increases to 1.9×1.9\times the success rate of the frozen model, while remaining comparable to internal block adaptation. We further demonstrate the same adaptation principle on a DINO-WM model for 3-D manipulation. These results suggest that effective test-time adaptation of world models does not necessarily require modifying their pretrained internal weights.
Nov 15, 2026

Where Should LLM Agents Adapt at Test Time? In Context, in Weights, or Both

LLM agents operating repeatedly in the same environment, such as a code repository, a database, or an application suite, encounter tasks that share latent structure. Experience can reveal this structure, allowing agents to reuse environment knowledge and learn from environmental feedback. Agents can adapt at test-time using this experience through two channels: retaining it in context and updating their weights. These channels are usually studied separately, leaving unclear what each contributes and when they complement each other. We compare in-context learning (ICL), reinforcement learning with verifiable rewards using GRPO, and self-distillation using SDPO, alone and in combination, using a shared online protocol and modular framework. Agents encounter matched task streams, are evaluated on each task before adapting to it, and learn only from their own attempts and environment feedback. We evaluate 14 adaptation mechanisms and a stateless baseline with Qwen3.5-9B on 109 tasks from four environments drawn from SWE-bench Verified, AppWorld, ScienceWorld, and BIRD-Critic-SQLite, using five task orders per environment. ICL provides the largest standalone gain: its best policy raises mean success across environments from 28.8% to 49.4%. Adding GRPO to this policy achieves the highest mean success, 51.6%. While adding GRPO improves the overall mean under every ICL policy, SDPO has mixed effects. Trace analyses show fewer lookups and shorter attempts under ICL, consistent with reusing knowledge of the shared structure. However, ICL also frequently skips inspection and verification for the current task. Adding GRPO to ICL helps restore these behaviors while preserving its reduced environment lookups. These observations suggest that reinforcement learning with verifiable rewards can refine how agents use retained experience, motivating the joint study of context management and weight updates.