Is an AI agent's answer all that counts, or should we also grade the route it took to get there?
What dimensions should trajectory-level scoring capture beyond final correctness?
This explores what a scoring system should measure about an agent's whole path to an answer (the steps, confidence, diversity, and process) rather than only whether the final answer was right.
This explores what a scoring system should measure about an agent's whole path to an answer, not just whether the last line was correct. The corpus points to several dimensions, and they are surprisingly different from one another. The first is process quality. Agent evaluation is moving from endpoints to full interaction sequences, and scoring is moving from correctness alone to process quality, recoverability, coordination, and robustness How should we evaluate agent behavior beyond final answers?. The case for it is strong. In one long-trace reasoning study, adding checks on intermediate states and policy compliance raised task success from 32% to 87%, because most failures were process violations rather than wrong answers Where do reasoning agents actually fail during long traces?. A final-answer score cannot see this kind of failure at all.
A second dimension is calibration. A binary right/wrong reward encourages confident guessing, because it never penalizes a confident wrong answer. Adding a Brier score, which punishes confident mistakes, as a second reward term improves accuracy and calibration together Does binary reward training hurt model calibration?. The same idea works at a finer grain. Confidence measured step by step catches reasoning breakdowns that a global average hides, and it lets you stop a doomed trace early Does step-level confidence outperform global averaging for trace filtering?. So a trajectory score can ask how sure the agent was at each step, as well as whether it ended up right.
A third dimension is what the trajectory is worth as part of a group or as training material. Sets of traces can be scored on collective usefulness instead of individual accuracy. Diverse mediocre traces can beat redundant strong ones, because an aggregator needs varied raw material to arbitrate Can diverse mediocre traces outperform redundant expert traces?. Successes and failures also deserve different treatment. Keeping only clean successes teaches models to tolerate errors, so one approach filters positives for quality and keeps diverse failures as negative signal Why do correct code trajectories teach models to tolerate errors?. Another treats successes as concrete demonstrations and failures as abstracted lessons Should successful and failed episodes be processed differently?. Here the score says what to learn from an episode, and not only whether it passed.
The fourth dimension is validity. Did the agent follow the intended path, or did it game the score? Reward hacking across weight training, output selection, and prompt revision shares one mechanism: optimizing against signals that only partly represent the real task Does reward hacking always stem from the same failure?. One response is to have benchmark operators issue claims of valid completion, backed by recorded infrastructure evidence and not only a terminal score Can infrastructure evidence replace terminal scores in benchmark validation?. There is a catch, though. Richer trajectory scoring does not remove the old evaluation problems of comparability, reproducibility, and mapping evidence to judgment. It moves them into a higher-dimensional space, so the field needs shared design protocols and not just a new format Do interactive evaluations actually solve the benchmark comparison problem?.
The corpus has less to say on whether the model's internal path can be scored directly. One note finds that reasoning tasks bend a transformer's representation path with 2-3x the curvature of lexical tasks, which hints that path geometry encodes difficulty Does transformer reasoning leave a geometric signature in representation space?. That note describes a measurement, not a scoring method, so treat it as a lead and not a recipe.
Sources 11 notes
Evaluation of agentic systems shifts evidence from final responses to full interaction sequences, and scoring procedure from correctness alone to process quality, recoverability, coordination, and robustness. This pattern appears across multiple agent benchmarks as a coherent design move.
Reliability for long-trace reasoning comes from checking intermediate states and policy compliance during generation, not from scoring final outputs. Adding intermediate verification raised task success from 32% to 87% because most failures are process violations, not wrong answers.
Binary correctness rewards incentivize high-confidence guessing because they don't penalize confident wrong answers. Adding the Brier score as a second reward term mathematically guarantees joint optimization of accuracy and calibration without trade-off.
Local step-level confidence catches reasoning breakdowns that global averaging masks and enables early stopping before traces complete. This approach achieves comparable accuracy gains to naive majority voting with far fewer generated traces, proving trace quality matters more than quantity.
SPIRAL shifts RL reward from individual traces to sampled sets, optimizing for complementarity rather than per-trace accuracy. Diverse mediocre traces outperform redundant strong ones because aggregators need raw material to arbitrate, not confirmation.
Show all 11 sources
GRPO-RoC filters positive trajectories for quality while preserving diverse failures as negative signal, allowing a 14B model to reach frontier math performance in 510 RL steps, surpassing much larger models with cleaner reasoning.
SkillRL demonstrates that treating successful episodes as concrete demonstrations and failures as abstracted lessons achieves state-of-the-art performance on complex tasks while using substantially less context than uniform approaches. The asymmetry mirrors human expert reasoning and avoids the degradation seen in uniform consolidation methods.
Reward hacking arises during weight training, output selection, and prompt revision through a shared failure: optimization against signals that incompletely represent the actual task. The substrate matters less than the misalignment between the scoring function and ground truth.
BenchShield enables benchmark operators to issue claims about valid task completion grounded in recorded infrastructure evidence rather than terminal scores alone. This shifts from a single number to a verifiable claim about whether an agent followed the intended evaluation path.
Interactive evaluation relocates core problems—comparability, reproducibility, evidence-to-judgment mapping—into higher-dimensional space rather than solving them. The field needs design protocols and shared standards, not format adoption, to make trajectory scoring interpretable.
Measuring intrinsic geometry across multiple models shows reasoning and analogy tasks carve paths with mean curvature of 0.71–0.83 rad, while lexical tasks produce only 0.27–0.31 rad, suggesting path geometry encodes task difficulty.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Local Coherence or Global Validity? Investigating RLVR Traces in Math Domains
- Reinforcement Learning with Rubric Anchors
- Beyond Semantics: The Unreasonable Effectiveness of Reasonless Intermediate Tokens
- What Characterizes Effective Reasoning? Revisiting Length, Review, and Structure of CoT
- BenchShield: Formal Model-Backed Instrumentation for Reward Integrity in LLM-Agent Evaluation Infrastructure
- Interactive Evaluation Requires a Design Science
- RLVMR: Reinforcement Learning with Verifiable Meta-Reasoning Rewards for Robust Long-Horizon Agents
- Diagnosing Harmful Continuation in Answer-Correct Long-CoT Training Traces