long-horizon-llm-agents-tool-use-terminal-benchmarks-token-level-confidence-predictive CLM-0602For 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.
Capability: Using the tools it is given
Observed on
Six open-weight models (26B-230B) on tau-2, Terminal-Bench-2 and GAIA; requires access to token probabilities and the full execution trace..
Sources
- Validated by: AUROC/AUPRC/best-F1 for trajectory failure prediction against entropy, sequence probability, SAUP, Tracer and UProp on three benchmarks and six models; average AUROC rose e.g. 0.608 to 0.656 (Qwen3.5-27B) and 0.694 to 0.718 (MiniMax-M2.7). Prefix-truncated trajectories show earlier detection, and selecting the lowest-risk sampled action raised task accuracy (Terminal-Bench-2 Qwen3.5-27B 0.105 random to 0.213). Ablations: removing graph modeling drops AUROC to 0.678; [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 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.Saying it does not know · reviewed by AI
- For agentic tool calls where a parameter is fabricated rather than grounded in the dialogue, sampling-based uncertainty detectors (token entropy, semantic entropy, SEP) fail at near-chance because the model fabricates plausible parameters with high confidence, while the model's hidden states at the moment the parameter is generated do carry a separable grounding signal.Stating false facts confidently · 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
- 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
- A linear probe on a tool-calling model's hidden state at the last generated token detects incorrect tool-calls — including wrong-but-well-typed argument values that parsers and logs do not catch — and works better on larger models and at middle-to-late layers than at the final layer.Using the tools it is given · 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: low. Falsifier as drafted: A trajectory-level uncertainty score built from local step confidence aggregated without dependency structure matching or beating relation-aware propagation on the same benchmarks and models, or the random-graph ablation performing as well as the typed graph. Automatic check flagged: figures not in the source: 0.016, 048.