data annotation teams

Abstract: Weakly-supervised instance segmentation (WSIS) based on either box-, point-, or image-level annotations has been proposed to overcome the labor-intensive annotation process in fully-supervised instance segmentation. Despite decent results, most WSIS methods only focus on a single annotation level, ignoring that some objects are easy to spot from point annotations, while others require box or even mask annotations. Therefore, we propose the Mixed-Supervised Instance Segmentation (MixSIS) task, which allows a flexible mix of annotation levels (image, point, box, and mask) within a dataset. MixSIS has a high practical relevance, as it allows optimizing the annotation process by annotating challenging objects in more detail, while saving time on simpler instances with less detailed annotations. To address MixSIS, we propose AnnotationsMixer (AnnsMixer) as a baseline. AnnsMixer is a truly mixed-supervised instance segmentation system, which generates pseudo mask annotations from all annotation levels, even mixed within a single image. We evaluate AnnsMixer on both WSIS and MixSIS tasks, largely outperforming state-of-the-art methods. In three case studies, we also demonstrate how the unique features of MixSIS and AnnsMixer can substantially reduce the annotation time (-90%) or improve the results (+34%) in instance segmentation with a simple annotation rule or leveraging heterogeneous data. Our code is available at https://github.com/chwilms/MixSIS

Abstract: LLM-based annotation has advanced rapidly, yet existing benchmarks remain misaligned with real-world practice. Actual annotation often involves lengthy, task-specific guidelines, and different annotators may produce divergent but reasonable judgments, especially on boundary cases, where single-labeled dataset fail to reflect this disagreement. This gap also limits existing single-agent and workflow-based annotation systems, which often lack adaptability to complex, evolving guidelines. To bridge this gap, we introduce ANT-A Benchmark, a real-world multi-annotator benchmark spanning 7 datasets across domains and modalities, with task-specific guidelines and 5 independent human labels per instance. ANT-A enables a distribution-aware evaluation, using Fleiss kappa to assess how closely model annotations align with human labeling distributions. We further propose ANT-OR, a multi-agent annotation framework that refines prompts by distilling contrastive decision rules from positive and negative feedback. Experiments show that ANT-OR improves agreement@>=2 accuracy by 4.5 points over the strongest baseline and achieves the highest alignment with human annotator distributions as measured by Fleiss kappa within 2 points. Code and benchmark are available at https://github.com/Krueske/ANT-OR.

Abstract: Stance classifiers make errors that can be costly to identify under limited human-review budgets. Prioritizing potentially incorrect predictions remains challenging, particularly across heterogeneous base models. We introduce B-CARE (Blind Classwise Assessment of Relative Evidence for Stance Error Ranking), a prediction-blind framework for identifying high-risk stance predictions. Given only a target and text, a fixed LLM verifier assigns ordinal evidence scores to all candidate stance labels without observing the base model's prediction or confidence. B-CARE constructs a comparative risk score by contrasting the strongest alternative-class evidence with the evidence supporting the predicted class. The resulting classwise evidence profile can be reused across multiple base models without training a separate error detector or calibrating risk scores. We evaluate B-CARE on four stance benchmarks and three heterogeneous base LLMs, covering 12 dataset-model settings and 5,206 unique test instances. Evaluation includes error-ranking discrimination (AUROC and AUPRC), selective prediction (E-AURC), and review effectiveness (ErrorCapture at fixed review budgets). B-CARE achieves a mean AUROC of 0.735 and a mean AUPRC of 0.524, capturing 41.06% of errors within the highest-risk 20% of predictions. Although B-CARE does not uniformly outperform strong confidence-based baselines, it provides a reusable, complementary evidence signal for prioritizing potentially incorrect stance predictions across heterogeneous models.

Abstract: Decision models return normalized probability distributions over candidate choices, which downstream applications can use to select among choices. Although their providers emphasize the reliability of these probabilities, calibration on tasks with a single correct choice does not establish their reliability when multiple choices are correct. We evaluate three decision models (Jev, Liquid d1, and GPT-6 Luna Decisions) on three diverse datasets (MetaTool, HotpotQA, and ECtHR) to study probability allocation and thresholding. Even when models rank every correct choice above every distractor, in 28% to 50% of those lists, at least one correct choice receives a probability below the uniform probability of 1/n, where n is the number of choices. We also present each model with five identical copies of a correct choice and observe probability disparities, with mean gaps ranging from 17 to 79 percentage points. We further show how these allocation disparities affect the recovery of multiple correct choices under probability thresholding. Decision-model probabilities can rank correct choices, but a probability threshold does not identify the set of correct choices.

Abstract: Linear relational embeddings (LREs) approximate a language model's subject-to-object mapping with an affine function. Their faithfulness varies enormously across relations, from near zero for person_father to the ceiling for country_capital_city. We investigate the factors that contribute to this variation and demonstrate that it can be significantly predicted from a readout that is not aware of the underlying relation. To do so, we propose a novel form of the Jacobian lens, the entity lens. Whereas the Jacobian lens is estimated across all tokens and concepts, the entity lens EL_t is trained to predict what the model will output for an entity of type t under the neutral prompt "{} is known for"; e.g., we train ELs for countries and persons. After estimating an entity lens, we measure how highly it ranks the model's answer for a relation without using any relation-specific information in estimation and prediction. Across 28 relations of the LRE dataset on Olmo-3-7B this relation-agnostic score correlates with LRE faithfulness at ρ= 0.7355 (p < .01). This suggests that linear decodability is largely a property of what a subject's representation already carries, rather than of the relation being decoded. Moving the unit of analysis from the relation to the individual fact, we conclude that new relation datasets and analysis methods are needed to address several problems in existing datasets, most prominently semantically weak first tokens (e.g., most predicted universities' first token is "University") and reference competition (e.g., the LLM confuses obscure actor names with competing non-actor entities that have the same last name). Our full pipeline is available at https://github.com/jonasrohw/entity-lens

Abstract: Complementary-label learning (CLL) trains a classifier from labels that name a class an instance does not belong to. Such labels are cheap to collect from crowdworkers or vision-language models (VLMs), making them attractive for large-scale annotation. However, both sources can mistakenly assign the true class as the complementary label. Such annotation errors are particularly harmful in CLL because the model is explicitly trained to suppress the correct answer. Optimal transport (OT) data valuation identifies harmful samples by comparing training data with a clean validation set, but transferring it directly to CLL creates a semantic mismatch: label $c$ means not class $c$ in the training but class $c$ in the validation. We show that this makes the OT label groups noisy mixtures whose behavior depends on the noise rate. The transferred score is informative only at sufficiently high noise, becomes uninformative at the crossover rate $1/M$ for $M$ classes, and reverses below it, where practical CLL annotation typically operates. We propose CAVA, which introduces a complementary label cost and a label margin that isolates the contribution of each complementary label to the transport potential. The margin keeps a consistent direction across noise rates, detects diagonal noise with a parameter-free threshold of closed-form precision, and allows detected samples to be recycled as ordinary-label supervision. On synthetic and human-annotated benchmarks, CAVA separates diagonal noise almost perfectly where directly transferred OT valuation performs worse than random. With only a fraction of the clean validation data, it also improves every CLL objective we evaluate, showing that OT-based valuation becomes reliable once the semantics of complementary supervision are modeled explicitly.

Abstract: This work presents a compact and reproducible pipeline for synthesizing and curating image-based instruction-response pairs for DCVLR. We draw from the Cauldron collection and use Magpie-style instruction and response synthesis with Qwen2.5-VL-7B Instruct to create concise, image-conditioned instruction pairs. The pipeline consists of instruction synthesis, conditioned on input images and controlled via a system prompt, sanitization, tagging, and filtering of the instructions based on the score obtained by a reward model (DeBERTa-v3). We retain samples with positive scores to maximize the quality of the generated instructions. We observe that score distributions have a median and mean reward score of ~ 1.53 and ~ 1.75 respectively, indicating that most generated instructions are of good quality. The resulting dataset contains diverse instructions in the math, spatial, and social reasoning domains.

Abstract: Target-aware multimodal instruction selection requires useful examples and an appropriate balance of supervision. Sample-level rankings can overrepresent similar content without optimizing this balance; group-level methods face coarse units, predefined set-level validation signals, and costly or inaccurate mixture search. We propose CLIVO, a target-aware framework connecting fine-grained semantic allocation, automatically learned validation objectives, and differentiable mixture search. Specifically, semantic clusters provide independently adjustable shares at the hundred-cluster scale. Correlations between per-example merged-proxy losses and real probe performance identify target-relevant validation examples; learned nonnegative weights combine their losses into an objective. A single backward pass through the merged proxy supplies gradients for all shares, avoiding an additional mixture-response regression model. Selecting 100K instructions from an 8M pool, CLIVO attains 97.2% of the full-data four-target average score with only 1.25% of the data. It improves on the strongest evaluated selection baseline by 1.16 points using random within-cluster sampling; combining learned quotas with sample ranking increases this advantage to 1.55 points. Analyses show that finer units benefit most targets, short-trained experts preserve mixture responses, sample weighting retains useful validation information, and direct search consistently reduces the learned objective. CLIVO also supports interpretable allocation, modular selection, and reuse across targets and granularities.

Abstract: Multimodal classification processes and relates information from multiple modalities to achieve more accurate predictions. However, existing methods typically rely on high-quality ground-truth labels, which are difficult to obtain in real-world scenarios. While sample-selection methods for learning with noisy labels aim to identify correctly labeled examples from noisy data, traditional methods primarily focus on unimodal settings and fail to exploit multimodal information fully. This motivates us to build a more reliable noise detector in multimodal learning. To this end, we theoretically analyze the relationship between representation structure and noise detection capability. Based on this analysis, we propose REFINE, which is a multimodal label-noise detection framework that jointly uses fused and unimodal representations for label-noise detection. Specifically, REFINE constructs discriminative eigenvectors through discriminative analysis of the target and background classes and selects trusted representation spaces with better noise detection capability for each class. Within each trusted space, REFINE measures the alignment between each instance representation and the discriminative eigenvectors. It then combines the subsets selected from these spaces. The combined set provides cleaner supervision for updating the multimodal classifier, thereby reducing the influence of mislabeled examples during training and improving model generalization. Extensive experiments across diverse tasks demonstrate REFINE's superiority compared to baseline methods. The source code will be publicly available.

Abstract: Large language models (LLMs) can solve many narrow tasks, but querying them separately for millions of related instances can be prohibitively expensive. Can LLM agents autonomously create cheaper solutions for such workloads? We call this ability "bottling": the ability to turn general capabilities into task-specific solutions that balance answer quality and amortised cost. We introduce BOTTLED, a benchmark in which agents receive an entire unlabelled workload and must complete it under fixed time, compute and LLM API budgets. Agents choose their own approach, such as training a small model or writing a reusable program. Across ten models and three tasks, we find that strong zero-shot task performance does not reliably translate into strong bottling capabilities. Models with similar zero-shot scores can differ substantially after bottling, and 48 of 60 bottling runs score below the lower bound of the 95% confidence interval of their model's zero-shot performance. Moreover, 31 of 60 runs underperform the stronger of two small-model distillation baselines with the same token budget. Nevertheless, bottling can yield substantial savings: on query-product relevance classification, Opus 5 retains about 82% of its zero-shot macro-F1 at roughly 657 times lower reported cost. Bottling is also competitive with Jev, a "system one" model built especially for cheap, repetitive inference: Opus 5 on the same task recovers about 94% of Jev's macro-F1 at a quarter of Jev's projected full-workload cost. BOTTLED provides a basis for evaluating and improving agents' ability to invest limited resources in reusable solutions for large, repetitive workloads.

Abstract: Multimodal large language models (MLLMs) advance rapidly, while conventional benchmark development lags behind, delaying investigation of newly observed capability gaps. Such investigation requires an expressive task format and an on-demand construction process: information-rich charts make chart question answering (Chart QA) suitable for probing coupled perception and reasoning. Automated Chart QA construction is intended to shorten the benchmark-development cycle by turning identified gaps into targeted samples on demand. Current methods, however, commonly separate target guidance from scratch generation: target-guided systems often require prepared data, charts, or templates, while scratch-generation systems primarily ensure artifact validity, without explicitly controlling whether newly synthesized requirements and content remain aligned with an externally specified diagnostic target. We introduce ChartBmkAgent, which turns an identified capability gap into targeted diagnostic evidence by constructing complete Chart QA samples from sparse error-taxonomy specifications. Throughout construction, a central harness governs specialized agents, requires stage-specific evidence of alignment with the original error category, and records the basis for each acceptance decision. On 300 taxonomy-wide samples, MLLM accuracies ranged from 32.7% to 84.3% with distinct category profiles, showing that generated samples reveal capability differences. Across three source-model comparisons, targeted follow-ups scored 50.0% versus 82.2% on matched controls ($p=8.96\times10^{-6}$); all six cross-model comparisons had the same direction, demonstrating targeted validation and diagnostic-data generation. Multiple evaluator models assessed whether each sample tested its specified error category; 86.4% met this criterion, providing empirical evidence of target preservation.

Abstract: Data quality now matters as much as compute for training language models. Much training data comes from human annotation of text, and interpretive annotation has no ground truth that could settle what is "accurate". Two lines of work respond to this. One combines annotators into a "ground truth" and measures how well they agree with each other; the other treats their disagreement as a signal. Both compare different people at one point in time. We measure something else: how well one reader reproduces their own reading of the same text over time. One expert human reader and three LLM families segmented three Sumerian myths and labelled the causal function of each segment with one of seven states. Across runs months apart, the human cut the text in much the same places but named the segments differently, in every myth. The models show no such consistent pattern: their gap between the two layers is positive in some myths and negative in others, and its size varies. The human's label changes are not random: the runs go through much the same functions but start them one step apart, while model runs start them at the same places. We argue that this pattern is a usable measure of data quality and a contamination check: a "human" annotation whose labels are as stable as its boundaries, and whose functions start in sync, looks like a model's.

Abstract: Modern machine learning depends heavily on massive datasets, but obtaining high-quality annotations at scale is often expensive. As a result, learning from noisily-labeled data has become common, making accurate estimation of the label-noise transition matrix crucial. However, existing transition matrix estimators rely on the fragile estimation of class-posteriors and do not provide finite-sample performance guarantees. In this work, we propose a novel methodology to estimate the transition matrix based on one-sided selective classification. This approach bypasses class-posterior estimation, provides finite-sample performance guarantees, and leverages flexible learning methods for binary classification. Moreover, we introduce effective algorithms to implement the proposed methodology and provide their refined finite-sample performance bounds.

Abstract: Structured span annotations, such as quantities with their units, uncertainty modifiers, and event classes, are expensive to create and hard to keep trustworthy once language models enter the loop. We present QuanReview, an open-source system for auditing and correcting such annotation layers. QuanReview aligns two annotation streams over the same documents at character level, resolves unambiguous cases by an explicit and logged policy, and routes candidate conflicts to a browser-based adjudication interface where reviewers accept either side, build field-level hybrids, or flag items for re-annotation. A campaign manager assigns documents to multiple annotators with configurable redundancy, computes agreement at document and span level, auto-merges unanimous documents, and exports the corrected layer in the original file format, so that it can replace the original annotation files directly. Applied to a 4,457-record humanitarian benchmark and an LLM extraction stream, the system fully auto-merged 8% of documents, applied automatic policy decisions to a further 1,513 records, and concentrated human attention on 3,131 candidate conflicts, a mean of 5.4 per reviewed document.

Abstract: With limited annotation budgets, choosing which images to label determines how much a model improves. Data-selection methods that use features from a separately trained model, or scene descriptions written by vision-language models, have been successful, but those signals do not directly capture changes in the model being improved. The target model's own internal features reflect what it has learned so far and change with retraining, making them a natural cue for choosing the next training data. However, feature rarity alone does not reveal the errors that matter for performance. Here we link internal features to prediction errors and their expected impact on performance and select images for labeling and retraining without using labels for candidate images. We evaluated the method with an object detector on two datasets and two pairs of random seeds. Adding internal features improved the identification of prediction errors in 15 of 16 conditions. When performance was averaged over successive labeling rounds, the method outperformed selection based only on feature rarity in all four evaluation settings and ranked among the top two of six methods. With other conditions held fixed, performance after retraining was again higher than with rarity-based selection, even though the latter collected more errors. With longer retraining, the proposed method ranked first among six methods. These results suggest that linking a model's internal features to its errors and their effects on performance may help select training images that improve performance, thereby allowing the model's current state to guide which images are labeled next.

Abstract: Reliable visual question answering requires correct answers when evidence is sufficient and abstention when it is not. We introduce a benchmark that connects complete-question evaluation with explicit evidence for its labels across PlotQA charts, CLEVR rendered scenes, and GQA photographs. Each question groups original and edited images, presented independently; success requires every supported answer and every required abstention to be correct. For chart missing-information labels, executable witnesses establish that admissible complete charts give different answers but identical pixels after masking. Scene labels follow source programs and edits, with a residual-cue analysis for photographs. Across 72,000 responses from six model configurations, the highest observed complete task success rates are 57.0%, 43.5%, and 33.7%, respectively. On charts, the strongest configuration achieves 96.2% per-view decision accuracy, yet 265 of its 835 groups with every decision correct still contain incorrect answers. Evaluating supported answers and necessary abstentions together exposes failures that answerability decisions alone conceal.

Abstract: An agent harness controls how a language model accesses information, uses tools, preserves memory, and checks its work. Improving this software is costly when each evaluation requires a long interaction with an environment. We introduce Vestrum, a framework that turns failures in execution traces into scoped harness changes without training the task model. Its organizing overhypothesis is that tasks of a shared kind may exhibit recurring failures whose remedies transfer within that kind. Vestrum expresses failures as recognizable classes, proposes changes across verification, retrieval, decomposition, and knowledge synthesis, and screens their scope before evaluating them as a bundle. A persistent lessons file informs subsequent proposals. Across five settings and two baseline harnesses, the frozen harnesses improve held-out performance: UltraHorizon rises from 47.6 to 59.8 over GAM, Terminal-Bench 4 Hard from 63.7% to 70.3% of checks passed over Claude Code on eight held-out tasks at 1.03x test cost, and cell-type annotation agreement from 67.5% to 77.8% on held-out sections of one slide, alongside gains on LoCoMo and AMA-Bench. Across our searches, verification grounded in evidence helped both intermediate steps and final answers, at lower cost at intermediate steps, while critics asked to rebuild finished answers broke more than they repaired. On the three memory benchmarks, Vestrum also scores above the evaluated GEPA configurations in every paired evaluation.

Abstract: LLMs are effective representations that assimilate vast amounts of knowledge during pretraining, but post-training is necessary for models to reliably access this knowledge and "know what they know." We observe an interesting rank equilibrium between knowledge stored in the weights and the data stream passing through the model. Across all model layers, we find that the hidden states (data stream) follow a U-shaped pattern, showing substantial compression in early layers and a steep rise during the late-layer decoding phase. In contrast, the weight rank follows an inverted U-shaped pattern, with very low rank in the early and late layers and high rank in the middle. We interpret this as an in-and-out information interplay: intermediate activations do not need to carry content that the weights can supply later, so they primarily preserve what the weights cannot provide. Motivated by this observation, we propose a model-aware data selection method, CAP (Counterfactual Assimilation Profile), which can determine whether a data candidate contains information accessible to the current model by utilizing the divergence gap in early- and late-layer representations between model-generated and reference responses. Across math, code, and science domains, CAP delivers 35.4% greater average improvement over the base model than the strongest baseline under different selection budgets. With only 10% of the data pool, CAP surpasses or matches full-pool training on math and science. We further show that CAP transfers to multimodal data selection and is robust to response horizon and noise.

Abstract: Validating an LLM-as-a-judge requires estimating its agreement with humans, yet annotation budgets rarely allow every item to be multiply labeled. We prove that this \emph{overlap sparsity} is the first-order determinant of wrong deployment decisions: at 5\% pairwise overlap, wrong-decision rates reach 25\% and the probability of selecting the wrong best judge among ten candidates is 65\%. The two actionable levers are overlap \emph{quantity} and \emph{allocation}. For quantity, we derive a minimum-overlap formula showing $ρ\geq 0.25$ suffices for non-borderline judges while borderline cases remain fundamentally hard. For allocation, a zero-cost stratified scheme halves false-rejection rates relative to random sampling when strata are informative. We validate on 10 LLM judges across four evaluation matrices spanning visual assessment, causal reasoning, and summarization.

Abstract: Fine-grained object detectors are commonly evaluated with confusion matrices, which show where the model is confused but not why, nor whether the confusion can be reduced. We argue that confusion can be attributed to distinct, separable sources, each quantitatively measurable, turning a passive measurement into actionable guidance. We present $A^2E^2$, a diagnostic tool that decomposes the sources of confusion along two axes, $\{$aleatoric, epistemic$\} \times \{$within-class, between-class$\}$, giving a $2\times2$ taxonomy that enumerates the source types. Each quadrant is measured by its own quantity, computed in one of three places (input geometry, output-space disagreement, and the bias-parameter posterior), so the two epistemic sources are separated by construction rather than by an empirical correlation. On fine-grained aircraft detection, the four quadrants become four named sources with their own remedy verdict: affinity (geometric similarity, irreducible from size alone), heterogeneity (geometrically heterogeneous sub-variants, pointing to re-labeling rather than more data), contested (an insufficiently trained but learnable boundary, improvable), and collapsed (a class starved of data, reducible). After attributing the confusion to a specific reducible source, we apply a targeted intervention and verify experimentally that it reduces the diagnosed source specifically while leaving the irreducible sources unchanged. $A^2E^2$ thus turns confusion measurement into a concrete, validatable and actionable "diagnosis" in which the same off-diagonal mass can carry opposite causes and opposite remedies. We also state this framework's limits, including which sources are only partially identifiable on this specific dataset and why.

Abstract: Large-scale text annotation brings expert insight to millions of documents, often through a codebook that AI annotators follow. Developing a robust codebook, however, takes months. Large language models (LLMs) could speed this process by applying an early codebook to the data, surfacing cases with strong LLM disagreement, and eliciting expert feedback to address them. We examined three ways experts can provide feedback for LLM codebook revision: (i) editing LLM-generated revisions driven by cross-LLM disagreement (Codebook Verifying), (ii) answering questions about LLM disagreements (Question Answering), and (iii) labeling disagreement cases with rationales (Rationale Labeling). Experiments on thousands of tutoring-session transcripts show that Rationale Labeling yielded the highest LLM-labeling accuracy (64.9%) against expert labels, outperforming the expert-revised codebook (57.8%). The best Question Answering setting also outperformed it (60.5%). Our work shows that LLMs can be used to strategically target expert attention, shortening months of codebook revision to days without sacrificing labeling performance.

Abstract: Large language models are increasingly used to scale codebook-based annotation in scientific research, but existing workflows provide limited support for translating domain experts' codebooks into reliable, revisable, and auditable prompts. Prompts are often treated as fixed instructions and hidden from annotators, making it difficult for non-technical domain experts to diagnose and correct model behavior when outputs violate codebook guidelines. In this paper, we present CALICO, a human-centered, codebook-aligned annotation workflow that treats prompts as editable, versioned, and optimizable artifacts. CALICO integrates codebook parsing, prompt generation, result inspection, prompt versioning, natural language human feedback, and label-supervised prompt optimization through existing optimizers such as GEPA, MIPROv2, and OPRO, together with our reflection-based optimizer, ReflectAgent. Empirically, we evaluate CALICO on domain-specific AI-companion chatbot conversation codebooks. Across evaluated dimensions, CALICO improves mean held-out performance by +13.0 and +7.4 absolute points for two coders, respectively. A coder-specificity analysis further suggests that optimized prompts capture coder-specific interpretations rather than only generic codebook clarification. CALICO runs as a web application that takes users from raw codebook materials to inspectable, exportable labels; the website, codebase, and live demo are released at https://calico-annotation.github.io/ under the Apache 2.0 License.

Abstract: Human corrections identify editable spans, but the examples receiving corrections may come from a selective feedback channel. We analyze this interaction at a fixed model checkpoint by decomposing a localized gradient into edited and retained untouched components. Squared relative selection bias is a ratio of quadratics whose derivative has the sign of an explicit quadratic polynomial. Localization can increase, decrease, or nonmonotonically change this diagnostic; its direction depends on component biases and geometry. Oracle importance weighting recovers the population mean for each fixed localization objective, but these objectives have different targets. Against one common full-gradient target, we derive the finite-sample mean-squared error, an analytic optimal retention coefficient, and a fixed-clipping extension. Exact finite-population calculations and 10,000 Monte Carlo repetitions per sample size verify the identities and counterexamples. Public human-post-edit experiments use two translation directions and pretrained models, with declared synthetic selection. An English-German extension differentiates 73.89 million native parameters. In all three declared settings, hard localization has higher relative bias but lower absolute bias than full retention. Untouched-component biases are nonzero and selected component means have negative inner products, so the general criterion applies where the simple unbiased/aligned explanation fails. Output-bias diagnostics show the same endpoint ordering of relative bias across both directions, with one interior maximum. Mechanisms reuse each language's records and include a mixture; they are not independent replications. The evidence separates relative amplification from absolute gradient error and establishes estimation properties, without inferring translation-quality gains or identifying actual complaint propensities.

Abstract: Many ML datasets are constructed by running a detector, heuristic, or model over candidate pools; accepted items become labels. Dataset precision is then governed by true-positive prevalence in each pool via Bayes, not solely by detector quality. Using one instrument and period, we hold a detector-defined event dataset plus an independent official index labeling every detected item as real or phantom. One detector, three pools yield phantom rates 81.7%, 9.0%, and 0.0%. Transferring precision from the two high-rate pools to the low-rate pool predicts 0.955 versus measured 0.183, a +422% error; the Bayes expression predicts all three within 3.3%. The detected response curve is an exact convex combination of a true-event and a phantom component (residual 1.1e-16), with phantoms outnumbering true events 473 to 308, so contamination is a second signal with detector-inherited shape, not additive noise. Contamination direction depends on the estimator: on identical windows one statistic is diluted and another inflated because its denominator is also contaminated. A common normalization turns the estimator into a mean of ratios whose expectation need not exist; on the same 335 events it returns 0.40 where the well-defined estimator returns 0.10.

Abstract: Autonomous research agents are increasingly expected to search the literature, analyze experimental evidence, and generate scientific hypotheses. These capabilities require multi-step evidence grounded reasoning that progressively acquires, integrates, and verifies evidence before reaching a conclusion. Existing multimodal benchmarks, however, largely evaluate final-answer accuracy, leaving open whether predictions are actually supported by traceable scientific evidence. We introduce Sci-MMR, a benchmark for multi-step evidence-grounded scientific reasoning built on structured argument graphs linking scientific claims, citation-grounded knowledge, visual evidence, and supporting regions. Sci-MMR comprises 235 multi-hop reasoning tasks spanning four scientific disciplines, with an average of nine figure panels per task. Evaluating eight frontier multimodal models, we find that answer accuracy consistently exceeds complete-evidence recovery rate by more than 20%, revealing a substantial gap that answer-only evaluation is structurally unable to capture. Through controlled interventions, we identify two fundamental bottlenecks. First, evidence acquisition: models struggle to extract complete structured evidence from scientific figures, accounting for 57.2% of failures. While cropping tools yield modest gains (+4.5 points), providing gold evidence improves accuracy by up to 37.0 points, indicating difficulty in assembling complete multi-region evidence. Second, evidence integration: models struggle to translate available evidence into correct conclusions, accounting for 31.8% of failures, while even with gold evidence the strongest model achieves only 69.1% accuracy on the hardest tasks. These findings indicate that current answer-centric benchmarks substantially overestimate the evidence-grounded reasoning capabilities of multimodal research agents