automated testing teams

Abstract: AI-driven code generation is evolving from handling small, localized code-completion tasks toward supporting the implementation of full software features. To support this transition, this paper introduces PR4Code, a dataset of 4,508 Java and 8,831 Python curated Pull Requests (PRs) collected from GitHub, each enriched with metadata, commit histories, and detailed code changes. Unlike resources focused on isolated code fragments, PR4Code captures the feature implementation process, delivering a large and diverse collection of real-world development tasks that can be used to assess AI-driven solutions. Our analysis reveals substantial variability in PR structure, commit granularity, and textual content. In addition, we release scripts to regenerate and update the dataset, ensuring reproducibility and maintainability.

Abstract: Recursive self-improvement requires models not only to acquire useful training experience, but also to decide when their current experience has ceased to be useful. Existing methods often focus on what experience to generate next, while paying less attention to when fresh experience should be introduced. Many also construct training targets through self-consistency or majority voting, making supervision depend on the model's own uncertain predictions. We address these two questions separately: training dynamics determine when to acquire new experience, while the answers to new tasks remain independently checkable. A-RSI monitors reward variation and accuracy in training logs to refresh a task pool when its learning signal declines, without consulting validation or test scores. A static library of model-authored programs produces multimodal questions and derives their reference answers; we specify the task-generation contract, automatically verify chosen properties of the reward and refresh rules, and separately audit the generated answers. The training loop uses no external training examples or human-written answer labels. On the primary eleven-benchmark suite, A-RSI raises the performance average from 57.77 to 61.05 with 800 fresh tasks.

Abstract: For large language models (LLMs) and multimodal large language models (MLLMs), model-reported self-confidence is often miscalibrated, yet it may still be useful for efficient evaluation. At the same time, LLM-as-a-judge offers a scalable alternative to direct human auditing, but its practical benefit depends on calibration quality and inference cost. This paper develops a unified statistical framework for both evaluation routes under alternative sampling strategies. For direct auditing, confidence-stratified sampling is used to recover population accuracy from partially verified outputs without requiring calibrated confidence. Finite-sample uncertainty is characterized, and the analysis shows how highly uneven confidence strata affect audit efficiency and label allocation. For judge-assisted evaluation, the judge is modeled through sensitivity and false-positive rate, yielding an error-corrected estimator with calibration uncertainty. The two routes are then compared through human-label and total-cost efficiency boundaries. Experiments on MMLU-Pro and MMMU-Pro show that miscalibrated self-confidence can still support statistically valid stratified auditing. Under highly imbalanced confidence distributions, proportional allocation substantially reduces estimation error relative to equal-per-stratum allocation. For judge-assisted evaluation, only a small subset of target--verifier pairs satisfies the idealized human-label efficiency condition, and automated inference costs can further limit potential savings. These findings clarify the statistical and economic conditions under which confidence-guided auditing and automated verification are useful.

Abstract: Coding agents can adapt their harnesses, the software, configuration, and instructions used to run them, based on experience from earlier tasks, but whether these adaptations improve their behavior on later tasks remains unclear. We study \retro, a skill that asks an agent to review a completed session and propose harness changes, and ask what changes it proposes and whether implemented changes reduce repeated work on later tasks. We ran Claude Code on 30 SWE-bench Verified issues, qualitatively coded the 125 proposals from \retro, implemented 24, and evaluated 12 harness changes on later issues from the same repositories in fresh sessions with and without each change. Most proposals referred to an event in the session, but 18 of 43 lint-related proposals recommended a practice regardless of what had happened. Among the seven evaluations in which sessions without the change encountered the targeted problem, two changes eliminated the targeted action, one reduced its occurrence, four showed no reduction, and some changes added work on unrelated issues. Harness self-evolution requires more than generating and implementing changes: it needs ways to establish whether a change helps on later tasks and to revise or remove changes that do not.

Abstract: Formal specifications can make AI-generated code mechanically checkable. Because user intent is typically expressed in natural language rather than formal logic, specifications must first be derived from it, e.g., via LLM-based autoformalization. Yet autoformalized specifications are themselves untrusted, so a key challenge is validating whether they capture user intent without relying on a costly oracle such as human review. In this oracle-free setting, a natural strategy is to vote among model outputs, but because model errors are often correlated, agreement among outputs may reflect shared mistakes rather than correctness. Our key insight is that these errors can be decorrelated through representational diversity. We achieve this diversity with contrapuntal languages: pairs of task-specific formal languages, each designed to express the intended specification, yet admitting no structure-preserving translation between them. An LLM formalizes the intent independently in each language, and we accept a specification only if a solver proves the two formalizations equivalent. To construct such languages, we present a synthesis algorithm that reuses fragments of sampled seed specifications as syntactic sugar over a base specification language. We implement the approach for first-order logic specifications and evaluate it on HumanEval+ and LiveCodeBench using GPT-5.4 mini and Qwen3.6-Flash with and without thinking. Over the strongest baseline, our method increases the proportion of accepted specifications that are correct, i.e., complete and sound, by 12.0% on HumanEval+ and 42.9% on LiveCodeBench. The accepted specifications judged complete reject 98.6% of buggy program variants on HumanEval+ and 96.2% on LiveCodeBench, compared with 92.8% and 78.3% for the strongest baseline.

Abstract: Coding agents based on frozen language models can improve their performance by refining the external context they rely on, without updating model weights. We study recursive self-improvement through behavioral specifications: natural-language code documentation that describes the behavior of program entities and guides agents in generating correct code. We formulate a self-improvement loop in which an agent generates code from its current specifications, evaluates it with execution-verified tests, and refines the specifications based on observed failures. The improved specifications then guide subsequent generation, progressively improving the agent's understanding of the codebase. However, recursive refinement is challenging because program entities are interdependent. Improving the specification of one entity can inadvertently degrade the correctness of others, a phenomenon we call output coupling. This creates a whack-a-mole effect in which local improvements cause regressions elsewhere, preventing consistent progress. We introduce SpecForge, a multi-frontier tree search framework that addresses this challenge with three mechanisms: (i) Pareto-guided state selection preserves complementary specification states and prevents premature commitment; (ii) dependency-constrained entity selection prioritizes callee-before-caller refinement; and (iii) error-diversified expansion explores multiple refinements using different subsets of test failures. On DevEval+, SpecForge achieves the best performance across five backbone models, including a 94.9% solve rate with Claude-4.5-Sonnet and a 5.7-percentage-point average improvement over the strongest baseline. The improved specifications also make the agent better at tasks outside the loop: they transfer to cross-language translation and to new-feature implementation, where they outperform the benchmark-provided requirement documents.

Abstract: Modern reasoning CAPTCHAs require fine-grained perception, compositional relation reasoning, and grounded browser interaction. End-to-end vision-language model agents remain brittle in this setting because a missed object, incorrect count, or confused relation can propagate directly to the final action. To address this problem, we present CaptchaAgent, a structured visual reasoning framework for interactive CAPTCHA solving. Instead of predicting an action directly from a screenshot, CaptchaAgent first constructs a visual intermediate representation (V-IR) that explicitly records entities, attributes, and spatial or topological relations. A symbol-only ASCII protocol further requires the parser to commit to a concrete object set and count. Building on this global representation, localized evidence refinement revisits ambiguous regions at higher effective resolution, while counterfactual consistency verification tests whether the resulting action remains stable under task-preserving transformations. Across three proprietary backbones on Open CaptchaWorld, CaptchaAgent improves Pass@1 by 21.3-25.8 percentage points over a commercially available browser agent baseline evaluated with the same backbones. On the separate offline Oedipus-data benchmark, it further improves average Pass@1 by 9.2-12.3 percentage points over the Oedipus solver across three matched backbones. The intermediate traces show that exact counting, continuous geometry, and ordered execution remain major sources of automation difficulty for the evaluated agents, providing practical guidance for designing more resilient compositional CAPTCHAs.

Abstract: Many long-horizon agents compact their context on a global rule, usually a token budget, blind to what the agent was doing. We ask whether the agent's recent behaviour predicts when a compaction will hurt. TRACE's public corpus of 590 harness-triggered AppWorld compaction boundaries replays each boundary from a re-executed prefix state under the pre-compaction context and under the summary, and records the burden of the next actions: calls that error or repeat a call already made. We find that pre-boundary history predicts post-compaction harm only weakly. An internally prespecified contrast by prefix placement is a wide null, and the naive "has-written" label behind it turns out to measure trajectory phase. The best extension-protocol trigger reaches held-out AUROC 0.66 (0.64 on the replicate's own label) against a same-boundary replicate of 0.72; the best frozen, interpretable trigger avoids 21% of harmful (positive-burden) boundaries while keeping 84% of compaction opportunities, and exceeds the random-rule expectation on count but not on burden mass (a post hoc comparison). Whether the best trigger beats a token-budget rule at matched retention cannot be evaluated on the release. We state what corpora should ship to answer it.

Abstract: As coding agents take on long-horizon software evolution tasks spanning multiple files and stages, longer execution trajectories introduce two coupled challenges: (1) accumulated histories strain context budgets, and (2) repository changes can invalidate earlier execution evidence. Existing approaches address these challenges through techniques like larger context windows, compression, retrieval, or repository representations, but often fail to reconstruct a consistent task state after a context refresh or verify whether recalled evidence remains valid. Thus, we introduce MemTrace, a provenance-aware memory system that preserves execution history and aligns its reuse with the evolving task (e.g., iterative cross-file repair) and repository state. MemTrace stores history as immutable Memory Traces anchored to key information (e.g., files, symbols, tests), and organizes their execution order and dependencies in a Memory Trace Graph. When context is constrained, working memory retains only compact Memory Anchors, from which the agent can reconstruct the latest execution state and locate evidence relevant to its next action. Before restoring historical evidence, MemTrace checks its validity against the current repository state and retrieves only what the next action requires. Across three complementary long-horizon coding benchmarks, MemTrace consistently outperforms all fully evaluated baselines under the same backbone and harness, improving DeepSWE pass@1 by 21.2 points, SWE-EVO Resolved Rate by 4.4 points, and SWE-Milestone Score by 17.8 points under Codex CLI.

Abstract: Computer-use agents need to reliably ground action targets in complex desktop scenes, where multiple applications, overlapping windows, and visually similar controls compete for attention. Existing training data rarely pair such scenes with dense annotations or vary them in a controlled way. We introduce DeskForge, a controllable desktop environment that composes and explores real applications to generate large-scale supervision for computer-use agents. It varies application states, content, window layout, appearance, and resolution, and fuses screenshots, accessibility trees, and window geometry into dense element annotations while recording the outcome of each executed action. Using this environment, we construct DeskForge-1M, a corpus of 1.2M annotated desktop observations containing 159.7M element instances. We fine-tune four vision-language models on 200K grounding examples drawn from DeskForge-1M. All four improve across held-out desktop conditions and on all five external GUI grounding benchmarks; for Qwen3.5-4B, accuracy increases by 11.51 percentage points on ScreenSpot-Pro and 10.11 points on OSWorld-G. The gains also translate to long-horizon task completion: under a fixed planner, the fine-tuned action models solve more WebArena-Infinity and OpenApps tasks, with Qwen3.5-4B increasing from 31 to 50 of 119 tasks and from 3 to 15 of 100 tasks, respectively. These results show that controllable composition of real desktop environments provides a scalable source of supervision for improving both GUI grounding and long-horizon computer use. The framework code, the dataset, and the fine-tuned model are available from the project page: https://saidgurbuz.github.io/deskforge/

Abstract: Algorithmic mathematical reasoning requires reliable decomposition, computation, and aggregation. Final-answer rewards provide limited guidance on intermediate errors, while successful execution does not guarantee mathematical correctness. This work proposes Function-Structured Graph Reinforcement Learning (FSG-RL), connecting subproblem graphs and Python implementations with multi-verifier feedback. The policy first learns to generate code from function graphs through supervised fine-tuning (SFT). Group Relative Policy Optimization (GRPO) then optimizes the policy using answer-gated rewards and span-level credit assignment. The framework also supports teacher supervision and structured memory. A benchmark curated from Grade School Math 8K (GSM8K), MathQA, MATH, and Omni-MATH pairs public function graphs with private verification specifications. Under a unified evaluation protocol, GRPO improves final-answer accuracy from 43.25% to 67.50% and full solution success from 32.25% to 52.25% over SFT. Continued reinforcement learning (RL) with teacher supervision yields additional gains. The gains extend beyond producing correctly formatted code, supporting verifier-guided reinforcement learning for mathematical reasoning. Code is available at https://github.com/ZihanLiummyycc/FSG-RL.

Abstract: Coding agents can return plausible patches that omit required behavior. These failures are hard to review because long traces and confident summaries often hide what was missed. We ask when a nominally weaker reviewer can reliably decide whether a patch solves its issue. We study 411 execution-labeled traces from three agents and 101 controlled cases. On 154 GPT-5.4 traces, structured but unchecked evidence raises both defect catch and over-rejection. We then provide official execution evidence as an upper-bound diagnostic. After choosing and freezing one of two formats per reviewer, five of six reviewers improve both rates on 122 held-out traces; two classify every trace correctly. Reviewer size is not a consistent predictor of quality. Because official tests are unavailable in deployment, we also evaluate a frozen cascade with patch-caused static errors and generated tests that first fail on the unpatched repository. On 121 scored held-out GPT-5.4 traces and 59 Gemini traces, its coverage is 0.89 and 0.86, risk is 0.33 and 0.26, catch is 0.76 and 0.80, and over-rejection is 0.66 and 0.67. Most false rejections occur when unresolved cases reach the reviewer. Official execution evidence shows the potential of weak review when decisive checks are available. Producing equally reliable checks without official tests remains the main bottleneck.

Abstract: GUI agents need long-horizon visual reasoning: they must interpret a changing interface while keeping a multi-step plan viable as earlier actions constrain later ones. Existing benchmarks evaluate grounding, computer use, and game play, but rarely test whether agents stay coherent across long chains of coupled decisions. Long-horizon visual puzzles expose this capability directly: a legal move that looks like progress can make the puzzle unsolvable, and the loss shows only several moves later. We introduce LongPuzzleBench, 114 levels in six puzzle games played through native GUI actions, where one objective can take a human over a thousand actions on persistent boards and dead ends go unannounced. With Native GUI Actions alone, the strongest agents solve most objectives, but success falls sharply on harder, longer boards: seven of ten general-purpose agents solve nothing harder than Medium, and none completes Bolt Unscrew Hard, which a human solves along with every other objective. Code Execution CUA does not close this gap, and its scores mix visual solving with algorithmic search. Controlled diagnostics trace these failures to one limitation that neither rules, state hints, nor failure memory removes: agents judge each move by the visible progress it makes, not by the future options it leaves.

Abstract: LLMs can generate programs for ARC-AGI-2 tasks, but the provided compute only allows a small number of attempts to generate, debug and validate solutions. Genetic algorithms can search and test many more programs, but random search rarely starts in a useful neighborhood of the solution space. We combine the two methods through a compact domain specific language (DSL). First, a quantized Qwen3.5-4B LLM generates an initial set of programs for each ARCAGI-2 task. Then, we use those programs to seed an initial population of starting programs, and use genetic algorithms to evolve these programs towards a solution to the given task. The DSL is designed such that every mutated program remains valid and can be executed. The initial programs proposed by the LLM solve 2 (3.3%) of the first 60 tasks of the ARC-2 public evaluation set. The genetic algorithm solves an additional 4, giving 6 correct test outputs in total (10.0%). If we try using evolving solutions without this LLM seeding, we do not arrive at any solutions at all. The results show that genetic search can improve programs generated by LLMs and produce additional correct solutions.

Abstract: Deep neural networks are vulnerable to adversarial attacks. Among black-box attacks, targeted decision-based attacks are particularly difficult: the attacker observes only the target model's top-1 label and aims to make it predict a prespecified target class under a bounded perturbation. Before perturbation refinement, the attacker must discover a direction that reaches the prescribed target region. This initialization step can incur substantial query cost. We propose GuidedRay, a targeted decision-based attack based on diversity-guided direction discovery. GuidedRay builds on two observations: target-class reference samples provide useful target-conditioned direction priors, and diverse candidates increase the probability of discovering a targeted adversarial direction. GuidedRay generates varied candidates from one or multiple target-class references and uses a one-query Fast Test to screen their induced sign directions. Once a feasible direction is found, GuidedRay applies Ray Search to reduce its decision-boundary radius. Experiments on CIFAR-10, CIFAR-100, and ImageNet demonstrate that GuidedRay consistently outperforms five state-of-the-art decision-based attacks at four evaluated query budgets from 500 to 5,000, with particularly pronounced gains in direction discovery during initialization. Against models protected by adversarial training or TRADES, it likewise achieves the highest attack success rate at all four query budgets.

Abstract: Agent benchmarks are substantially more costly to evaluate than conventional LLM benchmarks. Benchmark compression is therefore a natural solution, yet existing methods primarily model redundancy in task--model final-score distributions, which is important in agentic evaluation. To address this limitation, we analyze large-scale trajectories and identify six complementary process signals that are systematically associated with final agent performance. To disentangle agent performance redundancy from a complete perspective, we propose DualViewEval, an agent benchmark compression method that jointly exploits outcome and process relations to learn an exact-size miniset and predict the full-benchmark scores. Across five agent benchmarks and five representative baselines, DualViewEval achieves the best results in all datasets. With only 20 tasks, it achieves $24\times$--$40\times$ compression on APEX-Agents and BFCL, reducing mean absolute error (MAE) by $14.5\%$--$28.2\%$ over the strongest competitors while improving Kendall's $τ$ by up to $7.2\%$ relative to EssenceBench on SWE-bench Verified. The selected minisets further reveal capability differences among different agents, providing compact and diagnostic feedback for efficient agentic model development.

Abstract: LLM-based agents increasingly rely on persistent skills, i.e., reusable procedural prompts, to adapt without weight updates. Existing skill self-evolution methods directly revise skill text based on execution feedback, but each oracle evaluation requires a full agent rollout, creating a supervision bottleneck that confines search to failure-patching updates. Our key insight is that ranking is a smoother supervision target than absolute outcome regression: identifying which skill is better requires fewer oracle evaluations than predicting exact scores. Building on this insight, we propose SkillLift, which decouples skill search from oracle cost by learning an oracle-aligned rubric as a structured evaluation space. We formalize this as a bilevel optimization problem solved via alternating optimization: an inner loop uses the frozen rubric as a cheap surrogate to guide skill revision at no oracle cost, while an outer loop invokes a small number of oracle rollouts to re-align the rubric via rank correlation, amortizing oracle cost and stabilizing text-space updates. Experiments on complex agent task benchmarks show that our method outperforms existing auto-skill methods with 40--70\% less token cost compared to frontier evolving methods. Codes are available at https://github.com/WalteR-MittY-pro/SkillLift.

Abstract: Most of mathematical knowledge has been communicated through so-called informal use of mathematics and natural language. With large language models (LLMs) being highly adept in using natural language, they achieve strong performance, yet not perfect, in informal mathematical reasoning. Restraining LLMs to informal reasoning misses out on the opportunity to use the discrete verification abilities that machines offer through machine-checkable proofs. In this paper, we bridge the gap between informal and formal reasoning by integrating Lean signals into the informal reasoning process. We introduce Magenta, a training-free agentic pipeline that, given only a natural-language problem, produces an answer, expresses it as a Lean 4 statement, and constructs a machine-checked proof. A statement judge verifies whether the formalisation preserves the original problem, while an error-attribution judge routes failed attempts either to mathematical re-derivation or local Lean repair. Magenta achieves 100% accuracy across all evaluated olympiad benchmarks, including AIME 2025, AIME 2026, and HMMT February 2026. When paired with the open-weight K2-Horizon-7B reasoner, it solves all six IMO 2026 problems. Our analysis shows that statement adjudication is essential for preventing false certificates and that feedback-guided correction outperforms independent resampling on difficult problems.

Abstract: A model first sees an image from one physical measurement experiment, such as how far a block coasted, and must answer a question about a new trial, such as whether the block will pass a target after a fixed push. The initial experiment may provide enough information to answer, or the model may need another measurement, such as the object's mass, friction, restitution, or spring stiffness. We study whether vision language models can decide when to answer immediately and, when more evidence is needed, which experiment to perform. Current physical reasoning benchmarks usually evaluate only the final answer, so they do not directly measure this decision-making ability. We introduce a controlled evaluation where each problem provides one measurement image and four possible physical worlds created by combining two possible masses and two possible values of another relevant property. The model must either stop and answer or select the cheapest additional experiment that can resolve the question. We construct matched problem pairs where changing either the observed measurement or the question changes the optimal action. Since all possible worlds and experiment costs are known, we can explicitly determine the optimal choice. Across six open models and 144 physical parameter sets, direct responses repeat the same action for 95.1% to 100% of image pairs even when the correct action changes. Brief reasoning improves action switching, but the best model makes both decisions correctly for only 5.9% of image pairs. Additional analysis reveals failures in measurement interpretation, physical reasoning, and response formatting. By evaluating evidence selection separately from final answers, our benchmark reveals limitations in physical reasoning that conventional answer accuracy can overlook.