Training AI agents on outcomes alone is a known problem — and a new paper proposes a fix.
Researchers introduced TRIAGE, a credit assignment framework designed for agentic reinforcement learning, where a model takes sequences of real environment actions — searches, clicks, edits, object interactions — before receiving a final pass-or-fail signal. The standard approach, GRPO, spreads that outcome signal uniformly across every action in a successful or failed run. TRIAGE adds a semantic layer: a structured judge classifies each segment of a trajectory as decisive progress, useful exploration, no-progress infrastructure, or regression, then applies fixed role-conditioned rewards accordingly. The source of optimization direction stays with the verifier outcome; the role labels correct where credit lands within that outcome.
The distinction matters because outcome-only credit has two structural failure modes: it penalizes good exploratory steps inside failed runs, and it rewards redundant or backward steps inside successful ones. TRIAGE's authors show mathematically that role-conditioned credit is the optimal segment-level correction derivable from role labels, connecting it to lower-variance policy gradients — meaning the fix is not just empirical but theoretically grounded.
Across three benchmarks — ALFWorld, Search-QA, and WebShop — TRIAGE outperformed GRPO and two other baselines, with ablations pointing to regression detection inside successful rollouts as the dominant driver of gains. The efficiency improvement is notable: on completed ALFWorld and WebShop runs, TRIAGE reduced environment-facing turns by 10.4% and 14.8% relative to GRPO.
The hard part, as always, is the judge. TRIAGE's gains rest on reliable segment classification, and the paper's own framing — "whenever the judge is reliable" — quietly flags the ceiling.