partially-observable-text-environments-alfworld-scienceworld-letting-llm-agent CLM-0115In partially observable text environments (ALFWorld, ScienceWorld), letting an LLM agent query an externally maintained state estimate that keeps an explicit distribution over unobserved object locations raises task success more than querying a deterministic memory of what has already been observed, and the gain shrinks to near zero on a frontier model that already nearly saturates the benchmark.
Evidence for: Good regulator theorem (holds)
Capability: Tracking state through a long task
Observed on
Text-based household/science task environments where the only uncertainty is object location and the state space, prior and update rule are hand-specified; agents run under ReAct a.
Sources
- Measured success rate and steps on unseen splits of two text benchmarks, three models x two agent scaffolds, three runs (one for Sonnet). Belief-only beat base agent and beat a memory-only ablation on ALFWorld for Llama and Qwen; Sonnet showed little or no performance change. Belief accuracy is hand-engineered from the game engine's own priors, so the result isolates the value of the interface, not whether such a belief can be learned.
Disagreeing is the most useful thing you can do here. Nothing on this page waits for a reviewer: the counts above move when evidence lands, and every source stays attached with who brought it.
Related claims
- For LLM agents acting in partially observable environments, replacing the raw action-observation log in the prompt with an externally maintained Bayesian posterior over latent states makes the agent's action choice a function of the belief alone, removing surface-form sensitivity in which two histories with identical posteriors but different serialized text elicit different actions.Planning a sequence of actions before acting · reviewed by AI
- When a search agent is asked to forecast a future distribution rather than report the current state over the same cumulative literature archive, the forward-looking target changes what it retrieves — it allocates a smaller share of retrieved documents to the most recent window, which is the more predictive evidence — so carrying the agent's own state estimate forward beats its explicit forecast.Predicting future events · reviewed by AI
- For LLM web agents on cluttered pages containing many concurrent interface components, success falls sharply as component count grows, and having the agent first emit a localized sub-goal plus a task-relevant sub-observation before choosing an action recovers much of the loss, because self-assessment of the action is more reliable in the reduced context.Using the tools it is given · reviewed by AI
- For single-agent procedural tasks where the relevant state fits a fixed schema known in advance, replacing the appended conversation transcript with a structured execution state that the model rewrites each step — discarding the reasoning trace once the state patch is validated — holds accuracy flat as the horizon grows while history-appending runtimes degrade, and the gain comes from the structure rather than the shorter prompt.Completing tasks that take many steps or many hours · reviewed by AI
- For LLM agents on long-horizon web penetration testing, holding the backbone model fixed and moving global state out of the context window onto a persistent graph of verified facts and fact-conditioned intent edges raises task success most on the hardest tasks, where success depends on facts obtained dozens of steps earlier (88.2% overall and 75.0% hard, vs 44.1% and 25.0% for the strongest shared-plain-text-memory baseline).Completing tasks that take many steps or many hours · reviewed by AI
Notes
Drafted from the paper by a model and filed unreviewed. Visible here so it can be read, not because anyone has vouched for it: it does not move any technique's standing and does not count toward the internal scorecard. Drafted confidence: medium. Falsifier as drafted: An evaluation where the memory-only variant (observed locations, generic 'not yet observed' reply) matches the belief variant's success rate on the same tasks and models, or where belief access gives no gain over the base agent for small open models. Proposed technique, not catalogued: Expose a maintained belief distribution to the agent as a queryable interface.