llm-agents-whose-steps-condition-their-own-prior-outputs CLM-0588For LLM agents whose steps condition on their own prior outputs, per-step confidence that is calibrated against step-local success labels does not compose into a calibrated estimate of whether the whole trajectory succeeds, because correlated step errors make the product of step marginals diverge from joint success probability, and the divergence grows as prefix reliability falls over longer horizons.
Capability: Saying it does not know
Observed on
Holds where step labels are defined by local step goals rather than by survival (whether a successful continuation still exists); under a survival convention the non-composition di.
Sources
- Validated by: Computed (TC-ECE) on real agent traces from four models across three tasks up to a 50-step budget; pooling checkpoints hid overconfidence that was visible only at late horizons, and the agent's own stated confidence did not consistently beat a step-index baseline. No comparison against an alternative trajectory-calibration metric.. [truncated]
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 on dependent multi-step tasks, success decays geometrically with the number of steps under a single per-step reliability parameter that rises with model scale but stays below 1, so long horizons eventually collapse even for the strongest deployed systems — and the decay tracks step count, not context length, since bounding the context window steepens the decay rather than easing it.Completing tasks that take many steps or many hours · reviewed by AI
- For long-horizon LLM agents on tool-use and terminal benchmarks, token-level confidence (predictive entropy, sequence probability) barely separates successful from failed trajectories, and the separable signal lives in the dependency structure of the trajectory: scoring a trajectory by propagating per-step uncertainty over a graph of reasoning, tool-call and environment-feedback nodes raises failure-detection AUROC over both local-confidence and sequence-based agent UQ baselines.Using the tools it is given · reviewed by AI
- Having a model score its confidence step by step detects failures in multi-step reasoning better than one holistic confidence score for the whole answer, with up to 15% higher AUROC on two multi-step benchmarks, so where a chain fails is more visible per step than in the final answer.Carrying a chain of inference through to its conclusion · reviewed by AI
- When medical LLMs are given a clinical practice guideline in context and asked to produce the decision pathway a case requires, final-outcome accuracy does not track adherence to the guideline-defined step sequence: models can reach the right outcome while omitting, reordering, skipping or padding steps, and medically fine-tuned models score no better on pathway consistency than the general models they were tuned from.Competence in a specialist domain · reviewed by AI
- In multi-agent LLM pipelines over reasoning and code tasks with a pool of 3B-32B open-weight backbones, conditioning each per-step routing decision on a learned filtered memory rather than the full accumulated trajectory leaves accuracy about the same while cutting inference cost by roughly 40-50%, so most of the accumulated history is cost without accuracy value.Getting more from several models than from one · 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: Measuring prefix-coupling coefficients near zero across real agent trajectories, so the product of calibrated step confidences tracks observed trajectory success rate within sampling error at long horizons.