ai quality assurance teams

Abstract: Financial assistants answer questions that sound factual but depend on what the user leaves unsaid: what a data figure measures, which product variant is meant, and the investor's situation and the rules that apply. An assistant that ignores these can still answer with a real number from its data, one that answers a different question. We call this frame collapse. We study it in Indian mutual funds, where product variants and tax rules are formally defined, so the correct answer can be computed for every reading. One study tests each challenge. Given an ambiguously labelled figure, models used it, often under the user's label. Given every alternative labelled, they listed them all, but the layout decided which came first. Missing a fact about the investor, weaker agents assumed a default or an outdated tax rule; requiring the fact did not guarantee its use. Because such answers match their data, checks against the data pass them. An AI judge caught mislabelled figures only when told what each field means, and still passed answers that silently chose another reading. We show how these failures arise, not how often, and propose controls for financial assistants.

Abstract: Large language models (LLMs) are reported to be susceptible to reliability issues including hallucinations and reasoning errors. Existing quality detectors can predict unreliable outputs based on internal representations with decent accuracies. However, training these detectors often requires substantial domain-specific labeled data, which limits their applicability to new domains. We propose to address this challenge by transferring representational signatures across related domains. More specifically, we build a statistical framework that first extracts low-dimensional signatures for each domain, then uses cross-fitting to learn composition weights of domain-specific predictors. The weights adapt to each source domain's predictive relevance to the target domain. Our asymptotic analysis of a least-squares analogue characterizes the limiting prediction risk and aggregation weights in terms of domain discrepancies and representation bias. Experiments across multiple hallucination datasets and open-weight LLMs show that, with only 25 labeled questions in the new domain, adaptive transfer improves hallucination-detection AUROC by 6.8%. For reasoning-step prediction, it improves AUROC by 8.4%.

Abstract: Large Language Models (LLMs) increasingly serve as reasoning components in modern software systems, yet a model that answers a logical query correctly can contradict itself when the same problem is restated in an equivalent form. Static benchmarks evaluate queries in isolation and therefore fail to expose such inconsistencies. To address this limitation, we present LGMT, an open-source tool that detects reasoning inconsistencies through metamorphic relations grounded in first-order logic. Given a source test case in premise-conclusion form, LGMT generates logically equivalent follow-up cases and checks whether the model's entailment predictions agree across the pair, without ground-truth labels. The tool provides 20 built-in metamorphic relations, supports custom ones verified by the Z3 theorem prover, and runs from a CLI or a web studio, including an offline mock mode requiring no API key. Across six widely used LLMs on a bundled benchmark of 3,091 metamorphic groups, LGMT reveals inconsistency rates ranging from 20.84% to 29.44%. LGMT is publicly available on GitHub at https://github.com/MaxwelsDonc/lgmt-tool, with a demonstration video at https://youtu.be/gou4opptu1U.

Abstract: Checklist rubrics applied by large language model (LLM) judges have become a standard instrument for evaluating open-ended outputs and for rewarding them during training. Two assumptions surround this practice: that leniency or strictness is a property of the judge model, and that a rubric can be improved by asking a model to review it. We test both in a controlled grading task with planted flaws and on two human-labeled datasets, SummEdits and RAGTruth. Holding the judge and the outputs fixed, rubric content moves the pass rate of correct outputs from 2% to 100%, whereas changing the judge model or the rendering of the same checks moves it by a few points. Model self-review then drifts the rubric toward the reviser's own standard: it adds requirements the task never stated and deletes checks that human labels had taught, with three reviser models and on all three datasets. To measure and control this, we show that under all-must-pass aggregation the Shapley value of each check has an exact closed form, which yields per-check catch credit, false-fail blame and unique catches in linear time. Attribution shows that self-review deletes the checks that do unique work and keeps the checks that cause false fails. A change-control gate that combines these quantities with a provenance freeze on label-derived checks holds nearly every drifting revision chain at its starting quality, with identical decisions under three judge models.

Abstract: Rubric-based evaluation is widely used to assess large language model (LLM) outputs by decomposing quality into multiple criteria and assigning each a binary judgment from an LLM verifier. However, judgment certainty varies substantially across criteria. Existing methods uniformly aggregate these judgments into a final score, implicitly treating every criterion as equally certain and overlooking this heterogeneity. As a result, uncertain judgments remain incorporated into reward estimation without distinction, potentially yielding misleading training signals for downstream reinforcement learning (RL). To address this limitation, we propose Uncertainty-aware Rubric Weighting (URW), a training-free framework that quantifies criterion-level judgment uncertainty via conformal calibration. Specifically, URW elicits a soft uncertainty signal from a fixed pool of paraphrased judgment templates and converts it into a per-criterion weight using split-conformal $p$-values with provable coverage guarantees. Extensive experiments demonstrate that URW consistently improves pairwise preference accuracy and delivers stronger downstream RL alignment.

Abstract: Large language model (LLM) agents are increasingly used for long-horizon research and engineering, yet evaluating these capabilities remains expensive, time-consuming, and heavily dependent on human expertise. This motivates a natural question: can LLM agents themselves construct efficient benchmarks that preserve the conclusions of costly reference evaluations? We study this capability by formulating automatic benchmark development as a meta-evaluation problem, where an agent is given a task and an expensive reference benchmark and requested to construct a lower-cost proxy benchmark that faithfully distinguishes among candidate models and solutions. Based on this formulation, we introduce EvalBuilder, a framework for evaluating such benchmark-generation capabilities along three key dimensions: fidelity, efficiency, and robustness. Specifically, we assess whether generated benchmarks preserve model rankings, generalize to held-out models and freshly collected responses, and provide well-calibrated and discriminative scores while substantially reducing evaluation cost. We further provide an iterative environment in which agents can refine their benchmarks under a fixed time budget. Our framework provides a systematic testbed and benchmark for understanding when current LLM agents can develop useful evaluators and how benchmark construction itself can be increasingly automated.

Abstract: Verifiable instruction-following benchmarks emphasize output-constraint satisfaction but can overlook whether responses meaningfully answer the underlying request, creating an evaluation blind spot. We show that frontier models can receive full credit for responses with little substantive answer content. Using the same verifier-based criterion as a training reward can induce reward hacking, as publicly released checkpoints exhibit higher constraint satisfaction but lower response quality after reinforcement learning with verifiable rewards. To address this reward hacking, we introduce CALIF, a lightweight method that augments verifier rewards with a calibrated likelihood-based bonus to preserve response quality. CALIF evaluates response likelihood under the request alone using the existing frozen reference policy, requiring no additional judge or learned reward model. Across models ranging from 1.7B to 8B in reasoning and non-reasoning modes, CALIF improves constraint satisfaction on IFBench while maintaining response quality near or above pre-RL levels. It also outperforms corresponding base policies in 40 of 45 model-metric comparisons across three rubric-based instruction-following benchmarks, demonstrating gains beyond training verifiers.

Abstract: Verifiable instruction-following benchmarks often express each constraint through one fixed template. We test whether scores remain stable when the operational requirement is unchanged but its wording varies. We introduce WISE, a matched evaluation suite and reporting protocol instantiated on exact word count, keyword inclusion exactly once, and an inclusive 8--12 word range. Across 100 matched tasks, up to thirteen models from seven providers, and repeated generations scored over the complete visible output, wording alone produces substantial compliance shifts. In an avoidance-family panel, five avoidance and exclusion forms fall below the positive baseline, while constructional controls also shift compliance substantially: in the nine-model control panel, compliance is 54.9% for the original positive form, 48.2% for a longer positive form, 36.7% when the target appears later, and 33.8% for AVOID1. A strict JSON-structure probe shows wording sensitivity beyond counting, with a different direction of effect. Effect sizes, failure directions, weakest forms, and model rankings vary across realizations. Under the most disruptive exclusion form, the top-ranked model changes and 24.1% of strictly ordered model pairs reverse. Human validation further shows that unanimous agreement on an exact-count interpretation can coexist with substantially different model behavior. WISE supplements conventional scores with mean and worst-form compliance, wording gaps, failure profiles, and ranking stability.

Abstract: Multimodal large language models (MLLMs) are increasingly used as automated judges for instruction-based image editing and as reward signals for model training. However, systematically auditing whether these judges are influenced by cues irrelevant to editing quality is challenging because visual interventions may themselves alter the quality being evaluated. A judgment shift can therefore be attributed to bias only when the intervention is verified to preserve the underlying editing quality. To address this challenge, we introduce EditJudgeBias, a counterfactual benchmark with verified quality preservation, comprising 1,196 real editing samples and 13 cues injected across four evaluation sites. We verify quality preservation for the requested edit using calibrated multimodal validators, controls, and human inspection. We then audit five MLLM judges along three complementary dimensions: invariance to quality-preserving cues, agreement with human judgments, and stability of pairwise preferences. Importantly, observed shifts are evaluated against each judge's own zero-dose and re-query noise floors rather than against zero. Experiments show that quality-preserving cues move every judge beyond its own noise. Fabricated majority opinions increase ratings, irrelevant visual elements cause larger shifts than whole-image manipulations, and swapping candidate order reverses up to 60.9% of pairwise decisions. Edit-region cues also tend to reduce human agreement. The three measures characterize judges differently, showing that robustness cannot be captured by a single metric.

Abstract: Reinforcement learning with verifiable rewards (RLVR) has become an important approach for improving reasoning during post-training. Recent work suggests that some difficult prompts remain resistant to learning even when they occasionally produce correct solutions. We revisit this unlearnability phenomenon and find that the affected prompts do improve, at roughly one third of the learnable rate, while the difficulty-defined set used to study them is much less reproducible than expected. These difficulty labels are estimated from a limited number of sampled responses. Combining them across seeds can further change which prompts are selected instead of simply reducing measurement noise. We develop a sampling-based framework for quantifying this instability and determining how much evaluation is required for difficulty assignments to reproduce reliably. We also revisit the gradient-similarity evidence proposed to explain unlearnability and show that part of the observed separation arises because difficult prompts provide fewer correct rollouts from which their gradients can be estimated. Matching this sample count weakens the gradient difference but does not remove it. Overall, the slow-learning phenomenon survives our reanalysis, while both the prompts used to define it and the evidence used to explain it require more careful measurement.

Abstract: End-to-end accuracy on compositional tasks records how often MLLMs fail, but cannot distinguish whether a failure reflects an intrinsic deficit in the targeted capability or a cascading error from an upstream prerequisite. We propose a causal decomposition framework that isolates these two failure modes through controlled interventions on the prerequisite dependencies of each task. Our capability metrics (NC, IC, RC) score each task under unassisted, correct, or incorrect prerequisites to diagnose where failures arise; contribution metrics (N-Score, S-Score), adapted from probabilities of causation, quantify each prerequisite's necessity and sufficiency to determine why. We instantiate the framework in CADET, a diagnostic benchmark of 10 composite tasks decomposed into 46 unit tasks with over 33,000 human-annotated questions spanning perception, spatial, temporal, and cognitive categories. Diagnosing frontier MLLMs with our framework uncovers systematic patterns that end-to-end accuracy obscures. Capability-wise, supplying correct prerequisites eliminates 54\% of errors on cognitive tasks, lifting them from weakest to above spatial and temporal. Prerequisite-wise, causal contributions are concentrated in a few critical prerequisites, and supplying the single most important one alone captures 84\% of the gain from supplying all prerequisites.

Abstract: LLM assistants often produce more answers than humans can review before users see them. Most evaluations ask whether an answer is wrong, unsupported, or low-confidence. Bounded review budgets instead ask which answers should be checked first under a fixed review budget. Risk alone is not enough: a high-risk answer may be hard to repair, while a moderately risky answer may be directly correctable from available evidence. For generated-answer evaluation, we model review prioritization as exposure reduction, where review value combines estimated wrongness, intervention affordance, impact, and cost. We evaluate review queues with Wrong-Answer Exposure Ratio (WAER), the fraction of wrong answers left unreviewed, and post-repair residual exposure (PRRE), the fraction still exposed after deterministic benchmark-supported repairs. PRRE uses repairability rules that do not numerically reuse the affordance scores used for ranking. On a 720-item TAT-QA/SciFact stress benchmark, review-value ranking keeps answer-level WAER nearly unchanged at 20% budget (0.605 vs. 0.600) but lowers PRRE from 0.881 to 0.716. These results show that trustworthy LLM evaluation should measure not only error detection, but also how limited review capacity reduces exposed wrong answers.