Multimodal Large Language Models (MLLMs) process speech and text jointly, yet whether they exploit prosodic cues for pragmatic inference or rely on surface acoustic patterns has received little systematic investigation. We address this through sarcasm detection, evaluating Qwen2.5-Omni and Qwen3-Omni on Mandarin Chinese and English under five modality conditions that decompose the contributions of lexical content, vocal semantics, and prosodic structure. Adding audio systematically inflates false positives without improving true positive detection. Acoustic error diagnosis reveals that model errors cluster on a shared stereotype of expressive prosody, namely elevated pitch and irregular pausing, that diverges from the actual cues marking sarcasm in both languages. Targeted manipulation of only these two dimensions causally confirms the heuristic, inducing false positive rates of up to 60%. Applying the same manipulation template to Gemini~3 Flash Preview without modification replicates the effect, suggesting that the stereotype extends beyond the Qwen Omni family rather than arising from a single model architecture.
While the top model on OmniDocBench now reaches 96.34% overall on printed-document parsing, the ability of current models to handle challenging handwritten documents remains largely uncharacterized. Existing benchmarks focus on isolated text or formulas, overlook handwritten tables and real-world degradation, and report aggregate accuracy without explaining why models fail. We present WildHandBench, a benchmark containing 500 handwritten documents across three structures (free text, tables, formulas), four languages, and nine real-world scenarios. We introduce a Prior-Driven Error (PDE) metric that quantifies whether errors originate from language priors rather than visual evidence. Evaluating 18 state-of-the-art models together with calibrated human baselines, we find: (1) the best model achieves only 71.85% overall; (2) humans outperform all models yet the gap is narrow (77.09% vs. 71.85%); and (3) model errors are qualitatively different from human errors -- 63-91% of model errors are prior-driven versus only 49% for humans, exposing systematic reliance on language priors that conventional accuracy metrics cannot capture.
Sadab Shiper, Tawsif Tashwar Dipto, Mir Md Inzamam +1cs.CV cs.CL
In-the-wild Bengali scene text recognition is largely unmeasured: existing resources target handwritten documents or constrained sign-board parsing, report only aggregate edit-distance metrics, and evaluate either conventional OCR or VLMs, never both on the same in-the-wild data. To address this gap, we introduce BANGLAWILD, a benchmark of 2,535 Bengali scene text images, each paired with a verbatim gold transcription, two categorical axes, four diagnostic attributes, and an orthographically standard form where the in-image text deviates from canonical spelling. We evaluate fifteen VLMs and three conventional OCR systems under three prompting strategies, fine-tune 6 open-source models with LoRA, and complement edit-distance metrics with an LLM-as-a-Judge evaluation. Our results reveal a persistent gap in which larger models within the same family do not outperform smaller ones. Our fifteen-class error taxonomy shows that visual mis-recognition accounts for ~60% of errors in the strongest systems, while conjunct-related errors contribute under 2%, challenging a long-standing assumption in Bengali OCR research; the same visual dominant profile also holds across architectures, including the one conventional baseline that reads Bengali reliably. Prompt language mainly affects cross-script drift and LoRA reduces catastrophic failures in weak models without lifting the ceiling on already competent ones. Code and data will be publicly released.
Language models are widely used for generating and otherwise processing code (e.g., identifying code hallucinations, possible inputs, or predicting outputs); however, LLMs can make mistakes, which can be serious. One key issue is that models are trained on (still) largely human-written, and thus imperfect, code; it's not easy to find sufficiently large code corpora that are entirely free of bugs. Thus, other inference-time ways of reducing LLM errors, without additional training, are desirable. "Reasoning" or "thinking" modes, exposed as a togglable feature by hybrid reasoning models, do reduce errors; however, reasoning consumes additional resources. This paper asks if better performance can be achieved without always incurring the cost of reasoning. Human students of programming learn to avoid mistakes by (a) identifying them, (b) reflecting upon the cognitive lapses that led to them (essentially, "thinking through" the errors), (c) inferring general rules or lessons from these reflections, and (d) internalizing these lessons into rules. In tutorial sessions with an instructor, this is a common Socratic interaction. Examples of such internalizable rules might include the nugget "Before coding, restate the requirements to clarify them." Inspired by this process, this paper describes an approach where we first identify examples in which "thinking mode" in a (low-resource) LLM avoids errors. These errors, and their avoidance via "thinking" in the same LLM, are then examined by a bigger LLM to generate summary explanations; these are then summarized by a large LLM into brief advisory prompts. This approach works on many modest-sized models; in some cases, the "advisory prompts" thus learned can also be gainfully transferred to other models. We also present investigations into the nature of coding errors that language models make, and a characterization of when this approach can be helpful.
Post-training quantization to 4-bit weights is widely reported to be nearly lossless. We test this claim for multi-turn, tool-calling agents, where it now matters most. On $τ^2$-bench, across two open-weight model families in dense and MoE variants and two domains (eight cells, 456 episodes each, at 16-, 8-, and 4-bit weights), quantization indeed looks free on the standard metric. No cell shows a score change that survives multiple-comparison correction, and in the cell that carries the largest process damage, equivalence testing bounds the change within $\pm$7.5 points. The process tells a different story. Quantization amplifies the failure the model already exhibits at full precision (tool-name hallucination in telecom, with the same directional trend in retail entity errors) by up to 2.5$\times$ in volume (+17.6 points per task), while creating essentially no new failures. The failure set is the same at every precision (rank correlation $\geq$ 0.94, 0.18% novel events). The score stays flat because the benchmark's ten-error budget absorbs the extra failures. Shrinking the budget to two errors re-exposes a score gap of 17 points, and it does so only in the one cell where quantization added error volume, exactly as the masking account predicts. A targeted error-repair prompt, run for five telecom models at every precision, removes the damage exactly and only where it lives. Both diagnostics, the per-channel error rate and success under a shrinking budget, come from logs benchmarks already collect; we suggest reporting them alongside task reward.
Piyush Sao, Narasinga Miniskar, Pedro Valero-Lara +2cs.LG cs.IT math.NA
We study low-precision computation of C=AB with both factors quantized. We derive an exact finite-dimensional identity for the expected squared product error under independent, zero-mean entrywise errors with known variance fields; it holds exactly for non-overloading subtractive dither and for independent stochastic rounding, and we empirically assess deterministic round-to-nearest (RTN). Using the product-preserving equivalence AB=(AT)(T^{-1}B), we formulate contraction-gauge preconditioning: jointly choosing a factor representation and its sharing pattern before quantization. Preconditioning can reduce product error but may require extra transformed, quantized copies of the opposite operand: a shared transform needs one copy, a block-specific transform up to one per block. Within the bounded family of positive diagonal gauges (folds), a geometric program computes a globally optimal shared fold and a linear program decides whether the identity fold is already optimal. For other families we derive computable selection statistics -- tail index for scaling, profile spread for partitioning, coherence and weighted-Gram energy for rotations, slice-energy covariance for hierarchy depth -- with upper bounds for ranking heuristic candidates. Across twelve linear products from a trained three-block image classifier, median within-product rank correlations between dither-model predictions and deterministic-RTN errors are 0.937 at 8 bits and 0.918 at 4 bits. The GP fold cuts held-out product error over the identity fold by 18.0% (8-bit) and 20.5% (4-bit) in geometric mean, beats a SmoothQuant-style grid baseline at both precisions and on ten of twelve products, and lowers composed logit MSE by 15.4% and 26.4%. We thus provide exact stochastic product-error accounting, certified selection within the diagonal family, and a common objective for evaluating reusable transform candidates under RTN.
Tool-augmented language-model agents execute multi-step workflows over external systems, resolving an entity once and then acting on it across subsequent steps. Prior work shows that in single-step actions, agents select the correct tool but bind it to the wrong entity 24-26% of the time. We study what happens to entity bindings over time: do they stay correct, silently drift to a different entity, or, if wrong from the start, propagate and compound? We formalize binding drift (correct at step 1, wrong later) as distinct from error propagation (wrong at step 1, carried forward), and score them on disjoint workflow sets so the two cannot be conflated. In a controlled multi-step testbed (200 workflows, 580 entity-binding-scored steps, four enterprise domains, eight model backends spanning small to frontier), we find: (1) under controlled error injection, an entity lock (the intuitive "persist the first binding" fix) amplifies wrong actions from 907 to 2,746 (3.0x; bootstrap 95% CI [2.8, 3.3]), because it faithfully carries the seeded wrong entity into every later step; (2) the amplification reaches 8.5x on the most affected model (Claude Opus 4.5); (3) a practical LLM-based re-verifier (a single cheap second model call re-reading the original instruction) reduces wrong actions by 79% (0.21x; CI [0.18, 0.25]), closing the gap to within 1 percentage point of an oracle upper-bound (0.20x); and (4) in the natural (non-injected) setting, baseline agents drift on 18% of eligible workflows, with the per-step error rate rising across steps. Persistence and re-verification are not interchangeable: a defense that eliminates drift can worsen propagation, and a practical re-verifier nearly matches oracle recovery.
A trained flow or diffusion model is usually run with only a handful of solver steps, and the integration error this leaves behind is unevenly distributed across the image. We ask where that error is injected and how it reaches the endpoint, and answer with a signed source-and-transport accounting of few-step integration error, tested to first order. A perturbation experiment on five models at 256^2 resolution shows the learned dynamics spread local disturbances widely: near the start of sampling, under 10% of the summed endpoint response remains at the source. Signed one-step truncation residuals, propagated through the model's own linearized dynamics, reconstruct much of the endpoint error's direction and regional structure (cosine 0.81-0.87), and a region's error owes more to what arrives from elsewhere than to its own injection. Structure-destroying nulls, with protocols frozen before evaluation, locate what carries the account: randomizing contribution signs halves it, and reassigning which region receives each contribution, with content, norms, and signs intact, destroys it entirely. Where the injections land is readable from the model itself. The variation of its velocity or prediction field along the trajectory, a structure that emerges during training, predicts the final per-region gap (within-image rho of 0.57-0.70 on fine trajectories, weaker from the cheap solve alone). The prediction is partial because endpoint error depends not only on injected magnitude but on its sign, timing, and transport through the learned dynamics. A training penalty on the injected variation lowers few-step error, so the structure is one a model can be trained to change.
Most medical AI benchmarks measure whether a model knows the correct answer. MedFailBench asks a different question: which safety boundary failed? We present a clinician-built synthetic benchmark and failure atlas that labels medical AI errors by severity (1--5) and safety gate type (missed urgent escalation, unsafe remote dosing, unsafe discharge reassurance, evidence fabrication, unsafe protocol execution, source support gap). The current public release (v0.2.1) contains 44 clinician-reviewed synthetic cases with severity annotations, a live HuggingFace leaderboard preview, a safety gate taxonomy, a clinical severity rubric, and an automated pipeline for archiving model-response screening runs. No patient data, clinical validation claims, or model rankings are included. MedFailBench is released under Apache-2.0 and CC-BY-4.0 and carries the Zenodo DOI 10.5281/zenodo.21205535.
Vision language models (VLMs) have made remarkable progress in visual reasoning during the last decade. Most evaluations have used simple scenes (MS-COCO) that do not showcase complex human interactions or behaviors, only a handful of non-curated human descriptions as a benchmark, and have not focused on understanding the model's error types. Here, we introduce the Complex Social Behavior (CSB) dataset, containing 100 images depicting complex social interactions/behaviors. We analyze the progression of scene descriptions over a decade (2017-2025) of VLMs (four pre-Multimodal Large Language Models, MLLMs, and five MLLMs). We evaluate the accuracy of the models and 20 human descriptions relative to a gold standard on the CSB dataset and on a sample from MS-COCO. We analyzed five visual-cognitive error types: object detection, recognition, hallucination, scene understanding, and spatial dependence. The CSB dataset showed a more pronounced improvement than MS-COCO in scene description accuracy, with pre-MLLMs achieving much lower accuracy than the bottom-ranked human descriptions and MLLMs attaining accuracies similar to the top-ranked human descriptions. We show that MLLMs have eliminated the gap in scene description accuracy between simpler MS-COCO scenes and scenes depicting complex behaviors (CSB). MLLMs have almost eliminated all error types in our tested datasets, except for occasionally relying on different image regions for scene descriptions than humans do (spatial dependence error). We also show that detection, recognition, and hallucination errors have the highest impact on scene description accuracy. Together, our findings provide a more thorough evaluation of how visual language models have advanced over the last decade.
We study kernel-based operator learning in a two-stage sampling framework, where an offline kernel regression operator learns a discretized representation of the target operator from input-output pairs and an online kernel reconstruction operator recovers the output function from predicted observations. Our main theoretical contribution is an explicit budget allocation condition relating the number $N$ of training pairs, the number $n$ of input observations, and the output resolution $m$. The condition is derived from a coupled error analysis that interprets the surrogate as a reconstruction from approximate data. This yields a decomposition of the total error into reconstruction and learning contributions that can be analyzed independently. As a consequence, we obtain quantitative scaling laws describing how $N$, $n$, and $m$ must be coupled to guarantee convergence and to balance offline learning and online reconstruction errors. The resulting estimates extend previous analyses of kernel-based operator learning. We further introduce a physics-informed extension that incorporates knowledge of the underlying PDE at evaluation time. Rather than encoding constraints directly into the kernel, we augment the online reconstruction step by penalizing PDE residuals at collocation points. The method requires no retraining for new inputs. Numerical experiments illustrate the theoretical findings and demonstrate the effectiveness of the proposed physics-informed reconstruction strategy.
These notes give a proof-oriented introduction to diffusion models from the viewpoint of sampling, tracing a single arc from classical sampling dynamics to modern diffusion samplers, their error analysis, and inference-time control. Throughout, the material is layered into core definitions and identities proved in full, representative estimates proved under simplifying assumptions, and research-level theorems stated with a proof roadmap. The intended audience is beginning graduate students with a background in probability but no prior exposure to stochastic differential equations, stochastic numerics, or diffusion models.
Dyadic and circular convolution can both be computed in $O(N\log N)$ time using the Hadamard transform and the FFT-computed discrete Fourier transform (DFT), respectively. The Hadamard transform is preferable for its real-valued sign flips, yet its substitution for the DFT introduces algebraic error. We present three complementary results that characterize this error. First, we identify exact error cancellation: two input and two output positions are universally error-free, and no reordering of the output can eliminate this error. Second, the error operator is nearly full rank, while its null space has only logarithmic dimension. Third, the expected error is governed by a single alignment scalar, with a closed-form expression obtained by averaging over random filters. In general, the substitution error asymptotically doubles the output energy, except for filters in the universal zero-error subspace, which incur no error. Collectively, these results show that the substitution error is structured, predictable, and governed by alignment.
This paper develops an a posteriori error analysis framework for decoupled neural approximations of fully coupled forward--backward stochastic differential equations (FBSDEs). It provides an a posteriori error-analysis for the idealized discrete adapted trajectory. The main feature of the proposed formulation is the use of an auxiliary control process in the forward coefficients, which may differ from the backward component approximated by the neural network. This decoupling is useful in practical deep learning implementations, but it creates a control mismatch that must be included in the error analysis. We first establish a continuous-time stability estimate for fully coupled FBSDEs under perturbations of the drift, diffusion, generator, terminal condition, and auxiliary control input. We then transfer this estimate to the discrete-time setting and derive computable a posteriori error bounds depending only on the terminal defect, the pathwise residual, and the control mismatch. When the auxiliary control is identified with the backward approximation, the mismatch term vanishes and the bound reduces to the standard two-term form. Numerical experiments on a linear--quadratic FBSDE with an explicit reference solution and a multidimensional Burgers-type FBSDE without a reference solution illustrate the diagnostic role of the proposed indicators and the contribution of the mismatch penalty to the consistency and reproducibility of the numerical approximations.
Headline type-correctness (TC\%) of LLM autoformalization has climbed from $\sim$53\% to $\sim$76\% in two years, yet this scalar conceals which errors each method resolves. We propose a signal-coverage matrix that crosses the Lean elaborator (pass/fail) with a semantic-equivalence judgment (equivalent/not), sorting every output into one of four cells: true success (TS), type-only (TO), semantic-only (SO), or both fail (BF). On ProofNet\# and MiniF2F-test with DeepSeek V4-Pro across Vanilla, Lean-Retry, Sample-Filter, and Stratified Autoformalization (SAF): (1) the +34 to +36 TS gain across the three elab-feedback methods is $\sim$64\% type-stratum recovery, with SO flat on net (87.5\% of original semantic errors rescued, 8 newly created). (2) The TO-to-TS rate is 23/61 for each method (Wilson 95\% CI [26.6\%, 50.3\%]), and this stratum-level recovery rate predicts $Δ$TS on held-out methods to within 2/186 and renders $Δ$TC linear in the Vanilla elab-fail rate across six (model, dataset) cells ($R^2=0.96$). (3) The two judges disagree by 26 to 37 pp on elab-feedback outputs (vs. 7 pp on Vanilla), with 30 to 56\% of symbolic-judge false negatives traceable to elaborator-forced rewrites. The persistent residual reduces to two gold-formalization errors. TC\% gains should be credited by which cell moved, not by the scalar alone.
The trustworthiness of a retrieval-augmented generation (RAG) system depends on more than the answer it returns, yet many black-box uncertainty methods still read agreement among sampled answers as confidence. That inference fails when repeated samples condition on the same defective retrieval state. The state may be empty, with the model falling back on parametric memory, or populated by a coherent but wrong neighbourhood. In either case, the answers agree because the error is stable. The problem is recognised in deployed RAG, but it has lacked a name, a measurable signature, and a prevalence bound. We supply all three. We name the failure retrieval-state lock-in and diagnose it by separating the three objects a single confidence score conflates: the answer surface, the retrieved evidence, and the retrieval state itself. In an inspectable, ontology-guided knowledge-graph RAG (KG-RAG) system across six question-answering snapshots, we measure the agreement blind spot directly: at five samples per question, 42% of KG-RAG errors and 59% of dense-retrieval errors carry zero answer dispersion, so agreement has nothing to rank, while evidence- and retrieval-state checks still flag most of them. The decomposition supports an auditable decision rule: accepting an answer only when answer, evidence, and retrieval checks all agree that it is low-risk reaches 91.9% pooled precision against a 69.7% accept-all rate. The cost is coverage: it certifies only 7.7% of answers as low-risk. On the clinical calibration domain it reaches 100% precision under an automated judge; this is an in-domain automated-label upper bound, not a clinical safety claim, and still needs human validation. Confidence in RAG is object-specific: when answers agree, the useful question is which part of the pipeline to distrust.
Recent studies have shown that handwritten text recognition (HTR) systems perform worse on Arabic-script datasets than on Latin-script data. However, the reasons for this gap are still not well understood due to the lack of controlled comparisons. In this work, we present a comprehensive study of Arabic and Latin scripts HTR using a unified CRNN model for line-level HTR across nine datasets (including KHATT (Arabic), Muharaf (Arabic), NUST-UHWR (Urdu), PHTD (Persian), IAM (English), READ-2016 (German), and others) and di ferent training sizes (K in {100, 500, 1000, 2000, ..., Kfull}). Our results show the performance gap remains: it is large in low-resource settings, decreases with more data, but remains even at full scale, with a consistent difference of 5-7 CER points. We show that annotation quality matters, as many datasets contain labeling errors. Cleaning reduces error rates and narrows the gap, but does not eliminate it. In addition, we find that a fixed number of training samples provides less effective coverage in Arabic due to higher visual variability, requiring more data to learn similar representations. We compare recognition across datasets in terms of the number of text lines and the number of characters, showing an equivalence trade-off. We compare character frequency distributions across scripts and show that Arabic is significantly more heavy-tailed than Latin. Our error analysis reveals that around 30 percent of substitution errors in Arabic datasets (e.g., KHATT) are caused by confusion between visually similar characters, compared to about 15 percent in Latin-script datasets such as IAM.
Mika Mäntylä, Patricia Matsubara, Katia Romero Felizardo +5cs.SE cs.AI
Several studies have examined the use of large language models (LLMs) for title-abstract screening in systematic reviews (SRs), reporting mixed accuracy. However, questions of reliability remain largely unaddressed. In this study, we go beyond quantitative LLM-human agreement metrics and qualitatively investigate how and why LLMs fail. We also propose actionable recommendations. We analyzed disagreements between LLMs and researchers across six software engineering SRs and over 1,000 primary study papers. For each SR, papers were screened independently by human experts and LLMs in zero-shot mode, resulting in Kappa values ranging from 0.52 to 0.77. Qualitative analysis suggests that human-LLM disagreement results from recurring, identifiable causes, such as boundary ambiguity in key terms, keyword overemphasization, and incorrect topic inference. Based on these findings, we propose recommendations such as validating semantic understanding before deployment, running multiple LLMs, and focusing validation efforts on borderline cases. Future studies are needed to validate the impact of our recommendations, and community efforts are needed to develop normative guidelines on LLM usage in SRs.
Comparing text strings is crucial when evaluating and understanding the performance of various text processing tasks such as document recognition and audio transcription. With an increasingly complex landscape of AI-based handwritten text recognition (HTR), optical character recognition (OCR) and automatic speech recognition (ASR) models, there is a need for tools that facilitate evaluation in a flexible and reproducible way. This paper presents Stringalign, a Python library designed to simplify the evaluation process for automatic transcription projects and facilitate transparent evaluation. Stringalign's tools to examine and visualise both the rate of errors and the types of errors a model makes, give insights into possible improvements and help inform model selection for a particular task. Widely used string comparison metrics, such as the character and word error rates (CER and WER), although useful, can be ambiguous due to varying definitions of what constitutes a character and a word. Stringalign addresses this challenge by ensuring all preprocessing (i.e. normalisation and tokenisation) is transparent and easily replicable, and by providing tools to move beyond summary statistics and analyse common model errors. Moreover, Stringalign adheres to FAIR (Findable, Accessible, Interoperable, and Reusable) principles for research software while staying lightweight and easy to adapt into researchers existing workflows. In this paper, we discuss challenges with character and word level string comparisons and show through examples that where existing tools can yield opaque and sometimes confusing results, Stringalign provides an easy-to-use and unambiguous alternative.
Decision rules that enterprise experts apply tacitly -- in auditing, compliance, and contract review -- can be systematically recovered and improved through iterative error analysis. We present \textbf{Trace2Policy}, whose core mechanism -- \textbf{EISR} (\textbf{E}rror-driven \textbf{I}terative \textbf{S}kill \textbf{R}efinement) -- maintains a human-readable rule document as its optimization target: each round executes the rules on a validation set, clusters errors by root cause into MISSING, WRONG, or CONFLICT types, applies targeted patches, and commits only those that pass a regression gate. \textbf{For this class of compliance-sensitive, skewed-base-rate decision tasks, we identify rule quality -- not model capability -- as the dominant performance lever}: across five LLMs, one-shot distillation plateaus near $\sim$70\% on the deployed pool, while eight EISR rounds lift the same rules to 79.6\% when compiled into deterministic Python -- zero LLM calls at inference. \textbf{Execution form compounds the gain: in production, the same EISR-refined content runs 9.8~pp higher as compiled Python than as an LLM prompt, a form-and-engineering bundle the 22-day deployment matured together.} Deployed for 22 days at a major logistics carrier (3,349 audit cases), the compiled pipeline outperforms the pure-LLM baseline it replaced (72.7\%); on these calibrated, skewed-base-rate workloads, re-enabling LLM fallback monotonically degrades accuracy. An LLM-driven variant, \textbf{Auto-EISR}, reproduces this refinement at \$5--\$10 per cycle versus $\sim$70 expert-hours, and transfers to four public benchmarks spanning legal reasoning (LegalBench) and process-mining decisions (BPIC 2012) without re-engineering.
Large reasoning models (LRMs) often improve math and coding performance, but their effect on instruction following is unclear. We study IFEval with Qwen3 models (1.7B-32B), using same-weights Thinking ON/OFF controls; four Hunyuan models provide directional cross-family support. Aggregate pass-rate changes are small (-0.55 to -3.52 pp), yet 10-20% of prompts switch between pass and fail across modes, suggesting that thinking changes the pattern of errors--some prompts improve while others worsen--rather than uniformly degrading performance. Under a post-hoc Qwen3-derived grouping, constraint types separate into Planning (global counting, structure, coordination), which improves at the class level under thinking, and Precision (exact local form), which consistently worsens; the class-level Planning/Precision sign pattern holds directionally for all four Hunyuan models despite Hunyuan's opposite aggregate direction. Thinking also changes final-answer length; matched-length analyses substantially reduce the Precision drop, but a residual penalty remains. Analyzing thinking traces with a cross-encoder relevance metric reveals three patterns: Neutral shows a positive relevance-compliance link (r approximately 0.15); Planning shows near-zero predictive correlation (r approximately 0.02) despite measurable trace engagement, consistent with an execution gap between CE-measured trace relevance and final-answer compliance; Precision shows a small negative correlation (r approximately -0.05), with failing instances having higher mean relevance than passing ones. Activation patching across four model sizes (1.7B-14B) shows that Precision flip instances are more often restored than Planning flip instances (32-58% vs. 14-40% mean layer-restoration), with the largest gap at 14B (about 30 pp).
Useful audits reveal not only how often a model fails, but also where its failures concentrate. An auditor may test many candidate explanations: long inputs, indirect questions, distracting evidence, or combinations of these factors. The risk is selection. The largest observed effect may reflect a real failure mode, or it may simply be the best result among many tried. We introduce Janus, a procedure for deciding when a proposed error explanation is credible enough to report. The goal is not to generate new explanations, but to decide which ones hold up. The auditor starts with a fixed model, a labeled evaluation set, and a frozen list of candidate explanations, which we call descriptors. Janus scores each descriptor by its error-rate lift, then compares real descriptors with fake ones that have the same frequencies but are randomly assigned to examples. A descriptor is confirmed only if it beats this decoy floor on the data used for discovery and then repeats on separate held-out data. In a controlled audit of multi-table lookup tasks, Janus identifies the planted failure, confirming long-chain descriptors and their interactions. The LLM often stops partway through the lookup chain instead of reaching the final answer. On two public benchmarks, MuSiQue and LongBench v2, the SliceLine baseline flags plausible high-error pockets, but Janus confirms none of them. Ablations show why both safeguards matter. On LongBench v2, an uncalibrated fixed threshold reports 20 descriptors, the decoy floor leaves one, and the holdout check rejects the last one after its lift shrinks from 0.36 to 0.05. The resulting principle separates proposing explanations from reporting them. Candidates may come from any source, but only those that beat decoys and replicate on fresh data become audit findings.
In Video Instance Segmentation (VIS), classification, segmentation, and tracking objectives are jointly evaluated, but their individual contributions to performance loss remain opaque. We introduce a diagnostic framework that formulates identity and class assignment as an Integer Linear Program (ILP), yielding a model-agnostic oracle that hierarchically isolates each error source. Applied to seven VIS methods spanning online and offline paradigms across YouTube-VIS 2019/2021 and a diagnostic subset of OVIS, our analysis reveals a consistent picture. Tracking instability is a critical bottleneck for online methods, with gaps exceeding 20 AP under heavy occlusion, and grows sharply with video length and instance density. While semantic classification contributes meaningfully on standard benchmarks, its impact becomes negligible where tracking fails most. Although stronger backbones substantially lift default scores, they leave AP tracking gaps largely intact, confirming that temporal fragility is algorithmic rather than purely representational. To complement the oracle, we introduce TrackLens, a visual tool that translates gap magnitude into observable, query-level failure modes. Together, these tools provide a systematic foundation for targeting VIS's core challenge: robust long-term temporal association.
Erfan Loweimi, Zhengjun Yue, Andrea Carmantini +3cs.SD cs.AI cs.CL cs.LG
We analyse error patterns of raw waveform acoustic models on TIMIT phone recognition beyond the overall phone error rate (PER). PER is decomposed across three broad phonetic class (BPC) categorisations, and confusion matrices are constructed from substitution errors. Our models combine parametric (SincNet, Sinc2Net) or non-parametric CNNs with Bidirectional LSTMs, achieving 13.9%/15.3% PER on Dev/Test, the best reported results for raw waveform models on TIMIT. Transfer learning from WSJ reduces PER to 11.3%/12.3%, surpassing the Filterbank baseline. Per-BPC analysis reveals that BLSTM layers benefit transition-dependent classes most, while WSJ transfer learning improves consonants roughly three times more than vowels. Confusion patterns are consistent across raw waveform and Filterbank systems, indicating that the dominant confusions reflect inherent phonetic similarities.
Human-object interaction (HOI) recognition is critical for automatically analyzing student behavior in complex educational environments. Although state-of-the-art (SOTA) HOI detectors perform well on benchmark datasets, their performance often degrades when deployed in real-world training environments due to domain-specific objects, occlusions, and complex visual conditions. In this paper, we introduce a diagnosis-driven framework that integrates a triplet-level HOI error taxonomy with error-factor attribution analysis for real-world educational video data. We study this problem in the context of Critical Care Air Transport Team (CCATT) mixed-reality medical training. Based on an analysis of HOI failure modes and their causes, we develop a diagnosis-informed refinement strategy for adapting pretrained HOI models to the target domain. Experiments on the CCATT dataset show that this approach improves the macro-F1 score of a pretrained CDN model from 48.6 to 90.2 through targeted refinement guided by diagnosed error factors. These results highlight the value of detailed diagnostic analysis for informing targeted adaptation of HOI models in real-world educational environments.
Asaf Yehudai, Lilach Eden, Michal Shmueli-Scheuercs.CL cs.AI
Agentic systems are becoming more capable: agents define strategies, take actions, and interact with different environments. This autonomy poses serious challenges for overseeing and assessing agent behavior. Most current tools are limited, focusing on observability with basic evaluation capabilities or imposing static, hand-crafted error taxonomies that cannot adapt to new domains. To address this gap, we present Agentic CLEAR, an automatic, dynamic, and easy-to-use evaluation framework. It produces textual insights into the agent behavior on three levels of granularity: system, trace, and node. Agentic CLEAR operates above the observability layer, enabling seamless integration and featuring an intuitive UI that makes agent evaluation highly accessible. In our experiments on four benchmarks, seven agentic settings, and tens of thousands of LLM calls, we show that Agentic CLEAR produces high-quality, data-driven, insightful feedback. Our analysis shows strong alignment with human-annotated errors and the ability to predict task success rate.
We develop a learning-theoretic framework for understanding Chain of Thought (CoT). We model CoT as the interaction between an answer map and a chain rule that generates intermediate questions autoregressively, and define the reasoning risk of a hypothesis under this interaction. Our first result is a tight canonical decomposition of this risk into two terms with opposing roles: an oracle-trajectory risk (OTR), which captures the benefit of CoT and reduces to a target-domain risk in a domain adaptation problem, and a trajectory-mismatch risk (TMR), which captures the cost of CoT through error accumulation along mismatched reasoning trajectories. We then show that this cost is unavoidable without structure: if any one of the loss, the hypothesis answer map, or the chain rule lacks stability, the TMR can be arbitrarily large even when the OTR is zero and the hypothesis is uniformly close to the ground truth. Conversely, under stability, we prove a tight upper bound on the TMR governed by an exact amplification factor that identifies bounded, linear, and exponential error-growth regimes. Together, these results give a precise theory of when CoT helps, when it hurts, and what controls the transition between the two.
Automatic speech recognition replaces typing only when correction costs less than manual entry, a threshold determined by error types, not counts: fixing a misrecognized domain term costs far more than inserting a comma. Word error rate (WER) fails on two fronts: it collapses distinct error categories into a single scalar, and it structurally penalizes agglutinative languages where valid sandhi merges inflate scores. We introduce SCRIBE, a diagnostic framework that provides categorical error decomposition into lexical, punctuation, numeral, and domain-entity rates through sandhi-tolerant alignment with domain vocabulary injection. Human validation confirms SCRIBE aligns with expert judgment where WER does not. We release SCRIBE, an LLM curation pipeline, benchmarks, and open-weight rich transcription models for Hindi, Malayalam, and Kannada.
In recent years, Multimodal Large Language Models (MLLMs) have achieved remarkable progress on a wide range of multimodal benchmarks. Despite these advances, most existing benchmarks mainly focus on single-image or multi-image comprehension. In real-world scenarios such as document reading, information is often presented as interleaved multimodel contexts. This requires MLLMs not only to recognize the content of individual images, but also to identify relevant textual and visual evidence, establish fine-grained alignments between them, and reason over these aligned signals in interleaved contexts based on contextual evidence.However, there is still a lack of systematic benchmarks for quantifying the fine-grained understanding ability of MLLMs in interleaved image-text contexts. To fill this gap, we propose COHERENCE, a benchmark designed to evaluate the ability of MLLMs to recover fine-grained image-text correspondences in interleaved multimodal contexts. COHERENCE covers interleaved image-text content from four representative domains and contains 6,161 high-quality questions. Moreover, we perform a six-type error analysis, enabling fine-grained attribution of failures in interleaved image-text understanding to the specific capabilities missing in current MLLMs.