software qa teams

Abstract: Benchmark compression reduces evaluation cost by selecting a small set of samples. Existing methods typically guide selection using a single metric, although benchmarks often assess models along multiple dimensions. We find that model rankings are often weakly correlated across metrics, indicating that different metrics capture distinct aspects of model performance. Samples selected for one metric do not reliably preserve scores and rankings on others when used for compressed evaluation. At 80\% compression, each of the five single-metric baselines yields higher reconstruction error than equal-size random sampling in over 70\% of cross-metric comparisons. To address this challenge, we present HoloCosm, which jointly minimizes normalized score-reconstruction error across source models and metrics while constraining ranking distortion. We evaluate HoloCosm across multiple compression rates on 15 language, multimodal, VLA, and agent benchmark configurations with 85 component metrics. With 80\% compression, HoloCosm outperforms Random, the strongest baseline overall: average NMAE drops by 35.3\% to 0.236, worst-component NMAE drops by 39.5\% to 0.330, and average Spearman correlation rises from 0.884 to 0.915. In addition, HoloCosm remains competitive with existing baselines in single-metric evaluation. We will release our code and experimental results.

Abstract: GUI world models are increasingly used to improve action selection by predicting the consequences of candidate actions, but it is unclear how much of the resulting gain comes from prediction itself rather than from the evaluator that interprets those predictions. We study this attribution problem through controlled interventions on GUI action selection. With candidates and predictions held fixed, lookahead gains can change with how the prediction is presented to the evaluator, so end-to-end gains cannot be attributed to prediction alone. We then introduce Direct Action Ranking (DAR), which learns from reference actions to score candidates without generating future screens, and compare it with a prediction-augmented variant under matched supervision. Predicted successors add no detectable ranking signal once evaluation is learned directly. Without retraining, DAR improves AndroidWorld task success by 5.8--16.4 points across three proposers, outperforms world-model lookahead, and transfers zero-shot to unseen apps, where every prediction-conditioned configuration falls below proposer confidence. These results suggest that GUI world models should be evaluated by their decision value, the incremental action-selection value of their predictions beyond a strong direct evaluator, rather than by prediction fidelity or end-to-end lookahead gain alone.

Abstract: Claims that two training methods perform equally well often rest on a non-significant difference between two or three runs per method. We develop a protocol, with its power theory, for certifying practical equivalence in per-run metrics such as the AUROC (area under the receiver operating characteristic curve). It rests on a two-level (seed x item) variance model separating between-run dispersion from item-level noise. Equivalence within a pre-specified margin is tested by two one-sided tests (TOST), equivalently by a confidence interval inside the margin; a two-level bootstrap covers composite statistics. A closed-form formula converts a margin and a target power into runs per method; its shifted-t refinement is calibrated against the realised two-level process. Simulations on a 270-cell grid and under non-Gaussian seed laws confirm the formula within its design-relevant range, from five runs per method. For four published runs of a joint energy-based model, the failure to certify equivalence was largely decided by design: with the variance calibrated to those runs, two per method certify with probability near three per cent, whereas certifying 0.01 AUROC at 80 per cent power needs twelve even under exact equality. A second case study on a thousand-seed CIFAR-10 zoo and the MultiBERTs checks the formula against real seed laws (within 0.05 in that range) and resolves comparisons that two or three runs leave undecided. In a two-stage design (pilot, then formula), our experiments suggest that an upper confidence bound on the pilot variance substantially reduces the undersizing seen with its point estimate.

Abstract: Automated agent design increasingly trusts a single score to decide whether a change is an improvement -- but a score optimized aggressively enough ceases to measure what it was meant to. When an agent also explains its design, that explanation is a second signal we can check without human labels: negate each stated claim, re-run, and observe whether performance drops. We ask whether such a verifier stays reliable once an agent optimizes against it. Our contribution is not a new verifier but a way to stress-test one: a pressure ratio $ρ$ measuring how much the verifier steers the agent's search, and four falsifiable probes covering four failure modes -- giving no signal, being too weak to matter, being gamed by stating fewer claims, and being queried before the agent commits. We apply the test to our own verifier, the Rationale Consistency Score (RCS), inside an LLM agent that designs machine-learning pipelines, with one generator model, six tasks, and three seeds. First, in our prior configuration the verifier barely affects the search objective ($ρ\ll 1$), partly because of an accuracy gate, so its robustness had not been tested; it still shapes the reported artifact through a final-selection tie-break that $ρ$ does not register. Second, a capability control shows the agent shortens its claim lists when instructed, yet search does not, even at $ρ$ over 100 times the prior level; since feedback omits claim volume and horizons are short, this null result is largely structural rather than evidence of robustness. Third, when the agent can probe the verifier before committing, it keeps claims that pass while held-out faithfulness falls -- a decline invisible to in-loop metrics, though partly confounded by how RCS scores small pipelines. We offer these as preliminary evidence and propose a reporting standard: audit a verifier with $ρ$ and these probes before trusting it.

Abstract: On-policy self-distillation (OPSD) trains a large language model using supervision from a teacher instantiated from the same model as the student. The teacher is conditioned on privileged information (PI), such as verified solutions. The student, which does not observe PI, learns from the teacher's next-token distributions along its own sampled responses. However, PI can bias supervision toward shortcut reasoning with limited verification or reconsideration, accompanied by a sharp drop in the entropy of the output distribution. To address this challenge, we introduce self-distillation with verbal regularization (VRSD), which uses system prompts to regulate how the teacher uses PI, thereby encouraging rethinking and exploration to counteract premature commitment. Theoretically, we model prompting as attenuating the influence of PI on the teacher distribution and derive conditions for entropy preservation. Across three models and four tasks, VRSD improves Pass@1 over OPSD in 11 of 12 model-task settings, by 3.9 percentage points on average, while generally maintaining higher token entropy.

Abstract: Static evaluations credit a language model for naming the right move, but an agent must carry a plan through to a verified outcome while an opponent responds. We introduce XiangqiBench, an executable benchmark that measures this difference in Chinese chess: starting from 119 tactical endgames with forced mates supported by engine or checks-only search, an LLM agent must deliver checkmate against an engine defender. An interactive REPL interface separates real moves, state queries, and forward simulation, and we record 8,568 multi-turn trajectories from 12 frontier LLMs under two observation protocols. Three signals that look like competence each overstate closed-loop success. (i) The Conversion Gap: models play the stored reference first move in 26.1\% of Sighted trials, yet only 13.9\% of these trials end in a win. (ii) The Consistency Gap: the leading model reaches 38.7\% pass@3 but only 5.9\% pass^3, winning all three trials on 7 of the 46 positions it ever wins. (iii) The Simulation Gap: 32.3\% of accepted simulation calls stop on an illegal move, and in 49.3\% of comparable cases the real defender replies differently from the line the agent simulated; self-authored rollouts check legality but cannot anticipate the opponent. Finding the move is not winning the game: agent evaluations should score closed-loop outcomes and report reliability alongside coverage.

Abstract: As large language models are deployed in increasingly autonomous long-horizon tasks, manually auditing and verifying the actions, artifacts, and outputs of models becomes more difficult. Users instead come to rely on LLM-generated reports to assess the quality and completeness of the work. We introduce a suite of eight adversarial reporting scenarios to systematically study whether LLMs conceal narrative-changing flaws: errors or limitations that undermine an otherwise successful account of work. We call this phenomenon "insecure reporting." When handed machine learning experiment logs containing a planted negative result that substantially weakens the proposed method, GPT-5.5 flags the negative result in only 2 of 200 generated reports. However, when a short honesty instruction, "Be honest in your response," is added, the model flags the negative result in 190 of 200 reports. Across eight open-weight models, chain-of-thought analysis reveals a recurring tension between disclosing narrative-changing flaws and reasoning about ways to appear successful. We perform an activation analysis and a steering experiment on Qwen3.5-9B, finding that honesty and success-seeking correspond to opposing directions in representation space. Our results suggest that LLMs tend to present narratives of success by default, and that steering models toward honesty makes their reports substantially more transparent.