TraceProbe is presented, a trajectory-diagnostic framework that recovers what resolve rate hides and adds auditable diagnostic context to outcomes by localizing inspection targets, suggesting failure hypotheses, and prioritizing runs for review.
Abstract
Coding agents are ranked almost entirely by resolve rate: whether their final patch passes the target tests. Yet two agents can reach the same outcome through very different processes, and a single pass/fail label says nothing about why a run failed or why an accepted run spent extra steps, time, or tokens. This process evidence lives in the trajectory, which records a run's searches, reads, edits, tool calls, validation, and reversions. However, raw traces are heterogeneous and hard to compare across runs. We present TraceProbe, a trajectory-diagnostic framework that recovers what resolve rate hides. TraceProbe normalizes each raw run into a canonical nine-type action taxonomy with deterministic effect labels, then applies two rule-based modules: Insight names single-trajectory anti-patterns adapted from established debugging practice (e.g., search loops, verification skips), while Converge aligns pairs of runs and classifies where their behavior diverges under controlled references. Applying TraceProbe to 2,500 trajectories from five production settings on SWE-Bench Verified, we find that (i) file choice is too coarse to separate success from failure, whereas function selection and completion behavior localize it; (ii) Insight anti-patterns act mainly as corpus-level difficulty clues, with search loops the most stable; and (iii) even resolved runs differ in how quickly they reach relevant code and how much failed work they incur. Trajectory structure thus adds auditable diagnostic context to outcomes by localizing inspection targets, suggesting failure hypotheses, and prioritizing runs for review.
Outcome-only evaluation is the production default for LLM agents: show a judge the request and the final reply and ask whether it was handled well. The metric is structurally blind to an agent that reaches the right answer the wrong way. We measure that blind spot where ground truth is known by construction: a deterministic tool-using support-desk environment, a scripted oracle policy that always solves it, and a fault injector that breaks exactly one thing at a known step, stratifying faults by whether the customer-visible outcome survived (silent) or not (loud). Five judges (programmatic rules, outcome-only, step-rubric at two model sizes, and a self-consistency ensemble) are scored on detection, step localisation, fault typing, calibration, and cost over 400 trajectories. The outcome-only judge catches 84% of loud faults but 45% of silent ones while flagging 33% of correct trajectories; a step-rubric judge reaches 77% silent recall with zero false alarms at 3x the cost. No judge reads the final reply: an invented promise appended to an otherwise perfect trajectory evades the rules entirely and the step judge 82% of the time, and self-consistency triples cost while improving nothing. We argue that judge evaluations must stratify recall by outcome survival, and release the environment, the injector, all raw verdicts, and an analysis pipeline that rebuilds every number offline.
This work forks live SWE-bench agent trajectories at controlled points, rebuild the environment, continue each fork with a different model, and compare against same-model control forks that isolate sampling and replay noise, finding configuration-dependent temperature-0"determinism" is configuration-dependent.
Together, these results recast the evolved harness as a legible compensation layer, shaped jointly by the language's engineering demands and the model's behavioral gaps, rather than an opaque benchmark-tuned scaffold.
Siqi Yang, Qianlan Yang, Yu-Xiong Wang et al.· 2 citations
This work designs a contrastive study that combines controlled quantitative experiments with paired trajectory analysis and consolidates observations into a taxonomy of three high-level categories and twelve skill-use modes, showing that skills work when noisy trajectories become procedural anchors that stabilize execution.
Zhiyuan Jiang, Fan Huang, Hanwen Xing et al.· 0 citations
This work introduces Risa (Routing-Informed Steering and Arbitration): within trajectories, routing encourages diverse exploration and controlled convergence during patch commitment; across separately sampled trajectories, agreement at informative patch positions selects a final candidate.
Kang Chen, Junjie Nian, Yi-Xin Cao et al.· 0 citations
Agent evaluations report a tool-call rate read off the serving stack. That number can be zero while the model is emitting well-formed calls: the interface censors the trajectory before anything downstream sees it. On BFCL v4's own data, executor and scorer, holding weights, cases, decoding and seeds fixed and changing only the serving adapter, the same model scores 0.00 or 0.96 / 0.19. A 2x2 over chat template and parser locates the effect exactly: both main effects are exactly zero and all of it sits in the interaction -- no component is defective, and repairing one side of the contract buys precisely nothing. On tau-bench's 115 interactive retail tasks the same swap moves server-parsed calls from 0 to 636 and tasks reaching any tool execution from 0 to 103. Our probe reproduces the funnel across a 21x scale range of Qwen2.5-Coder: the server parses 0/100 at every size while well-formed emitted calls rise to 80/100 at 32B (~72 after calibration against an adjudicated gold standard). Under a matched envelope, across a comparable scale span, the silent fraction stays at 0-2, a prediction committed to the repository before the run. Llama-3.1-8B's 23% rate of calling the task function itself as a tool falls to 0 under one strict:true flag. The mismatch reaches inside the training loop, and its consequence is scale-dependent: in verl's AgentLoop at 7B, 45 of 115 generations carry a complete call; 0 are accepted, 0 execute, 0 return an observation. At 1.5B the same zero is over-determined, so we report the two scales separately. At evaluation time, repairing the adapter restores the mechanism but not a significant outcome gain: parsing 0->84, rescues 0->9, pass rate 53->62 (n.s.). We release a 98-line preflight check that catches every silent failure here. The observed tool-call rate is not a property of the model alone; it is a property of the model-interface stack that measures it.
Wen-Bo Wang· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
