AI
COUNTERMEM: World-Model Verified Counter-Factual Memory for Language Agents
Publication: cs.AI updates on arXiv.org
Author: Hongji Pu, Ruixiang Tang, Yongfeng Zhang
Source published: 29/09/2026, 04:00:00 UTC
Briefing created: 29/09/2026, 05:20:04 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 29/09/2026, 05:18:24 UTC.
Briefings
COUNTERMEM is a reinforcement-learning framework for constructing verified counterfactual memory in language agents. It evaluates local alternatives from a state copy or reset using executable world models like tests, proof checkers, and solvers. The system stores improvements with original and corrected actions, checked outcomes, and conditions for reuse. A learned memory-use policy balances task success and interaction cost while the base LLM remains fixed. Evaluation on 12 benchmarks across six domains shows COUNTERMEM improves ReAct and Reflexion by 12.6 percentage points on average. Task-run tokens decrease by 7.7-42.0% in four-domain comparisons, excluding offline selector-training costs. Verification and persistent storage weaken gains, while applying verified corrections to unsuitable decisions can reverse performance drops.
Key points
- COUNTERMEM evaluates local alternatives from a state copy or reset using executable world models such as tests, proof checkers, and solvers to construct verified counterfactual memory.
- The framework improves ReAct and Reflexion on 12 benchmarks across six domains by 12.6 percentage points on average, with task-run tokens decreasing by 7.7-42.0% in four-domain comparisons.
Context and limitations
Evaluation is limited to 12 benchmark settings across six domains; code will be released upon acceptance. Results exclude offline selector-training costs and may vary with different world models or task types.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 29/09/2026, 05:20:04 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeiarj6gtgjk6xfic4sol7av44cf3ulql5uc27nogtuf3qqp5tppaj4