Evals

User feedback helps models repair targeted faults, but LLM judges often miss those repairs. Agent evals should retain human feedback and avoid treating model preference as ground truth.

Cliff trains on the location of a rollout’s first reasoning error, rewarding the valid prefix and penalizing the suffix. It offers finer supervision without a specialized reward model.

A label-free Dutch-book test finds internally inconsistent probabilities from language models, especially when prompts add logical complexity or irrelevant context.

Mechanistic tests suggest two open LLM judges route local errors early and form ratings late, giving eval builders concrete layers and components to probe for judge failures.

CordisBench finds that models lose reliability as harness lifecycle interactions grow, while deterministic execution semantics can replace costly reasoning on controlled cases.

StudentSim turns sparse user histories into individualized simulators that model both current behavior and response to guidance, a pattern for testing adaptive agents before live deployment.

BLOOM-WILT makes rare model behaviors easier to surface by adapting both the auditor’s strategy and target decoding. Its results also show that elicitation method can reverse safety rankings.

S3Gym tests whether agents can explore, judge outcomes, and reuse experience. Raw history, compressed memory, and training each help in some tasks, but none is a reliable default.

Ironclad argues against optimizing token volume or spend alone: measure reviewed, merged code and fix review and CI bottlenecks before judging coding-agent ROI.

Formal verification gives agent-generated code a machine-checkable contract beyond tests. The hard part shifts upstream: humans must define and validate what correct means.

SWE-Prime finds that filtering coding-agent traces by process and segment quality can beat training on every resolved trajectory, reducing noisy imitation from redundant or risky steps.

RedEvoAgent turns prior jailbreak trajectories into a compact attack skill, then keeps only validated improvements. It offers a more interpretable way to probe tool-using agents for unsafe actions.

A multi-GPU kernel benchmark shows coding agents can improve with tools and iteration, but still struggle with communication scheduling, partitioning, and transfer choices.

Agentic checkout spans product feeds, tool discovery, checkout state, and delegated payment authority. Builders need protocol-specific adapters and evals because ACP and UCP have not converged.

TraceML exposes why coding agents stall on long ML work: they repeat narrow optimization loops instead of alternating work types, revisiting discarded ideas, and pivoting like experts.

Correct answers can hide invalid computation. Trace Integrity proposes executable, replayable contracts for data agents and reports 45.8–59.1% correct-answer/invalid-trace rates across three SQL modes.

Snowflake’s rollout favors narrow, high-accuracy coverage, staged adoption, and log-driven iteration over connecting every data source before launch.

Einstein Arena suggests multi-agent environments can outperform fixed workflows when they expose verifiers, shared solutions, forums, and incentives. Its results also show why benchmark shortcuts need active testing.

Plausible outputs can hide consequential omissions that generic LLM judges miss. Production evals need real failure discovery and retrieved expert judgments, not a frozen rubric alone.

Per-task model routing cut the demonstrated coding session’s cost from 44¢ to 14¢ with similar completion time, but builders still need workload-specific evals to validate quality.

A tool-exposed ontology steered models closer to human therapy patterns without fine-tuning, showing how explicit action vocabularies can improve agent behavior.

AI4AI-Bench tests whether agents can rewrite training algorithms, not merely tune runs. The best system scored 0.250, with most submissions never changing how the model learns.

Per-problem self-improvement claims can arise from inference and evaluation noise. This audit argues every transition statistic needs a measured null from frozen baseline replicates.

When an agent moves from fixed chains to model-led orchestration, memory, tools, and sandboxes, its old answer-quality evals cover only part of the system. Evals must follow the architecture.

When production data cannot be retained, generate eval cases backward from sampled labels and reasoning paths, build records in layers, and let domain experts steer the pipeline.

Vertical agents need narrow jobs, proprietary data, observability, and expert judgment. Generic models and self-grading cannot establish whether domain-specific output is actually useful.

Hippocratic AI’s voice stack uses specialist models, parallel checks, contextual speech recognition, and offline verification to avoid a single clinical-agent failure point.

Relationship agents can reinforce a user’s one-sided story unless clinical expertise, adversarial evals, risk escalation, and strict data handling are designed into the product.

A controlled study encoded authenticated internal-state evidence into unchanged answers, suggesting generated text could carry provenance signals, but not that current models reveal them naturally.

Tested compliance guards often ignored the governing rule and classified from scenario cues. Builders should counterfactually swap policies before trusting a detector as an audit control.

Static computer-use benchmarks can reward memorized action scripts rather than adaptation. Vary task state, verify every generated case, and calculate uncertainty across both actions and environments.

QuoteBench shows that shell-command scores can conceal failures introduced by serialization and reparsing. Agent evals should identify the execution path, not attribute every result to the model.

Cursor’s AIUC-1 certification combines a controls audit with adversarial testing of live agents. For enterprise evaluation, it adds behavioral evidence beyond conventional data-security attestations.

Agent quality problems often live in traces, not source code. Mine production runs for recurring failures, compaction drift, and counterfactual model choices, then validate changes with replayable evals.

Agent evals tied to a model, tool call, or harness can decay after an upgrade. Keep checks in code, prioritize production failures by onset and reach, and use agents to investigate detected anomalies.

A Sonar talk argues that coding-agent gains need repeatable verification, citing a three-month productivity spike alongside persistent static-analysis warnings and complexity.

A production refactor shows why coding-agent evaluations need acceptance criteria and end-to-end verification: fast output can still be incomplete scaffolding.

SkillProx evolves reusable agent instructions through measured retries, regression rollback, and utility-based pruning, reporting a 3-point accuracy gain over its strongest baseline.

P-Bench tests whether agents choose statistically valid methods, not merely execute code. Fisher-R1-14B improved single-trial success over DeepSeek-V4-Pro by 21% on average.

SABRE turns a Markdown test design into generated VLM stress tests, then filters and repairs candidates. It offers a repeatable pattern for refreshing evals as models improve.