From Sequence to Structure: Relational Uncertainty Propagation for LLM Agents
RUPA scores an agent’s whole execution path by tracking how uncertainty moves through its reasoning, tools, and feedback.
The paper argues that token-level or step-level confidence misses failures seeded earlier in a long agent trajectory. RUPA turns the run into a directed graph, then propagates uncertainty across temporal and semantic dependency edges. The authors report stronger uncertainty estimates, earlier failure detection, and better uncertainty-guided execution on τ-2, Terminal-Bench-2, and GAIA with six open-source LLMs. Source: HF Daily Papers' note.
The paper argues that token-level or step-level confidence misses failures seeded earlier in a long agent trajectory. RUPA turns the run into a directed graph, then propagates uncertainty across temporal and semantic dependency edges. The authors report stronger uncertainty estimates, earlier failure detection, and better uncertainty-guided execution on τ-2, Terminal-Bench-2, and GAIA with six open-source LLMs. Source: HF Daily Papers' note.
score 4