QA
-
Beyond Accuracy: Counterfactual Fragility and Demographic Bias in Clinical Evaluation of LLMs
2026-10-10 19:28:45Chaitali Deb Purkayastha, Bharath Kumar Bolla, Vishnu Surya Reddy NandiAbstract
Clinical LLM evaluation often emphasizes answer accuracy; however, accuracy alone does not test counterfactual consistency or demographic robustness. We evaluated six LLMs on 150 MedQA USMLE questions using two automated perturbation tests to assess their performance. The counterfactual validity (CFV) test asked each model to make a minimal, plausible clinical change that would make a different answer correct. The demographic robustness test added six demographic prefixes to the same vignette and compared the answers and explanations with a no demographic baseline. Of the 900 CFV attempts, 228 (25.3 %) were valid and 672 were invalid. Across 5,400 demographic comparisons, 1,097 answers were changed (20.3%). Automated judging identified 3,128 stereotype evidence flags, including 1,932 in the broad Other category. MedGemma 27B achieved the highest accuracy (87.1%) and CFV (63.3%), lowest answer change rate (16.0%), and low mean Explanation Demographic Dissonance (EDD) score (0.169). However, its accuracy still exceeded its CFV, indicating that correct answers do not guarantee reliable performance on the counterfactual validity task. OpenBioLLM had the highest answer change rate and EDD, whereas GLM had the highest stereotype flag rate. These findings show that accuracy, CFV, answer stability, EDD, and stereotype evidence capture different evaluation aspects. Because all judgments were automated and no clinician validation was available, the results support safety screening but do not establish clinical deployability of the model.
Abstract (translated)
URL
https://arxiv.org/abs/2609.32807
PDF
-
Right Answer, Wrong Reason: Accuracy, Consistency, and Consensus Are Misleading Indicators of LLM Faithfulness in Clinical Decision Support
2026-10-10 19:26:58Bharath Kumar Bolla, Chaitali Deb Purkayastha, Vishnu Surya Reddy NandiAbstract
Clinical Large Language Models (LLMs) achieve strong medical-exam accuracy; however, a correct answer does not guarantee that the explanation names the concepts that actually drove the decision. We introduce three lightweight, directly interpretable metrics for this faithfulness gap: the Explanation Stability Index (ESI), which measures reasoning consistency across repeated queries; the Causal Faithfulness Score (CFS), which tests whether cited concepts drive predictions via concept ablation; and the Perturbation Stability Score (PSS), which measures robustness to semantic-preserving paraphrases. By evaluating six LLMs on 150 MedQA-USMLE questions (900 model-question observations), we found that only 23.3% of the cited clinical concepts were causally necessary. Correct answers had lower CFS than incorrect answers (0.212 vs. 0.398), answer consistency negatively predicted CFS (Spearman r = -0.466), and model pairs could agree on answers while sharing only 8.8% of cited reasoning concepts. These results show that accuracy, consistency, and consensus are incomplete safety signals for clinical decision-making support. The evidence is behavioral rather than mechanistic: concept ablation tests counterfactual sensitivity of outputs, not internal circuits.
Abstract (translated)
URL
https://arxiv.org/abs/2609.32817
PDF
-
To Defer or Not to Defer? Online Learning with LLMs and Congested Experts
2026-10-10 19:14:51Vasileios Saketos, William Reveillard, Alexandre Proutiere, Sebastian Dalleiger, Ming XiaoAbstract
Modern AI systems increasingly combine large language models (LLMs) with human experts, using the LLM to handle routine queries while deferring uncertain cases for expert review. When expert capacity is limited, however, deferral decisions must account not only for the reliability of the LLM but also for congestion in the expert queue. We formulate this as an online sequential decision-making problem in which, for each query, the system observes the confidence signal of the LLM's answer and must decide whether to output the answer or defer the query to a capacity-constrained expert. The reliability of the confidence signal is initially unknown and must be learned from expert feedback, which is both selectively observed (only deferred queries are labeled) and delayed by queuing. We propose \texttt{KL-PBD}, a learning algorithm that uses Kullback--Leibler lower confidence bounds on the unknown reliability parameters to guide deferral decisions. For $K$ confidence levels and horizon $T$, \texttt{KL-PBD} achieves $O(K\log T)$ gap-dependent regret when the accuracy at every confidence level is bounded away from the deferral boundary, and $\widetilde{O}(\sqrt{KT})$ worst-case regret. We establish matching lower bounds, showing that these rates are optimal up to logarithmic factors. Finally, we validate our approach on synthetic instances and real-world question-answering datasets, namely GSM8K, NuminaMath, and MedMCQA, using Qwen LLMs.
Abstract (translated)
URL
https://arxiv.org/abs/2610.15128
PDF
-
Trace Evolution Bench: Measuring Whether Agents Improve from Their Own Environment Traces
2026-10-10 16:45:52Fangfei Li, Chenyang Zhao, Long Wang, Feng Tian, Lv GuoAbstract
Self-evolving agents are usually scored by endpoint Gain: held-out success after an update minus success before. That difference can credit a frozen rescore as evolution. We introduce Trace Evolution Bench (TEB), a protocol that separates that rescore from a context or weight update, together with three frozen 80/20/40 packs. An update is accepted only when validation does not fall, held-out identifiers never enter the update, and a text update is not ranked with a weight update. The primary scoreboard is coding, graded by hidden unit tests in a closed workspace. On that pack (40 held-out tasks, one family), DeepSeek v4 Pro under no_evolve never changes its policy yet moves 7.5, 5.0, and 12.5 percentage points across three seeds. Of nine text-update cells inspired by GEPA, ACE, and ExpeL, seven have positive excess above that seed's own band, and one gepa_llm seed moves down by the width of its band (-12.5, regression 0.48). GLM and MiniMax do not reproduce that seven-of-nine pattern. No method stays above its own band on every seed of all three models: GLM has two paired cells above their own bands, and MiniMax expel_llm seed 44 is a third (regression 0.30). The methods were fixed on a validation screen before held-out was read. Office, graded by rule-bounded delivery, and manufacturing table QA are separate packs: office DeepSeek does not exceed its own bands (widest 17.5 points), and manufacturing has no paired positive excess beyond its seed band.
Abstract (translated)
URL
https://arxiv.org/abs/2610.15055
PDF
-
Litmus: Zero-Label, Code-Driven Metric Specification for Evaluating AI Systems
2026-10-10 12:50:27Prajjwal Gupta, Prasang Gupta, Vishal Bhutani, Apoorva Sharma, Sumanth Chundru, Waqar Sarguroh, Kevin PaulAbstract
As agentic LLM systems move from prototypes to deployment, evaluating them has become more important and more difficult. Evaluation goals are often left implicit. Without a clear account of what a system is expected to do, how it can fail, and which failures matter, metric choices become difficult to justify, interpret, or validate. We present Litmus, a zero-label system that designs evaluation and monitoring metrics for AI pipelines by eliciting evaluation intent from source code and targeted interrogation. Litmus first identifies what must be measured and why, then converts those answers into constraints for a justified, per-stage metric portfolio. We evaluate Litmus on three real, code-defined pipelines against AutoMetrics and three DynamicRubric baselines. Litmus achieves broader concern coverage on account grouping at matched size, spans more stages, produces a near-zero-redundancy portfolio, and, under a zero-label design-time constraint, ranks first in validity against per-row quality labels on all three pipelines, decisively on scientific QA ($\rho{=}0.72$ vs. $0.47$ for the strongest baseline on the stitched single-vector score), with overlapping intervals on the other two. Results support a shift from automatic metric implementation to automatic metric specification: before asking which metric to compute, evaluation systems should ask what must be measured and why.
Abstract (translated)
URL
https://arxiv.org/abs/2606.23403
PDF
-
Co-adaptive Rule Routing and Constraining for Structured Knowledge Retrieval
2026-10-10 01:45:48Zhisheng Qi, Haoyu Han, Jiaxi Li, Bing He, Huiyuan Chen, Franck Dernoncourt, Ryan A. Rossi, Mahantesh M. Halappanavar, Jin Lu, Yu Zhang, Tianming Liu, Yu WangAbstract
Structured knowledge retrieval acquires task-relevant evidence from structured knowledge sources for downstream tasks. Existing methods rely on unconstrained retrieval or predefined query-rule supervision, such as author-paper rules for scientific discovery and anatomy-drug rules for precision medicine. Such supervision is rarely available in practice. We propose Co-Adaptive Rule Routing and Constraining Retrieval (CoRE), which jointly learns which reasoning rule to assign to each query and how to retrieve evidence under that rule, without predefined query-rule pairs. A routing agent selects a rule, and its corresponding retriever traverses the knowledge base and ranks evidence that satisfies the rule's constraints. Because routing and retrieval are interdependent, CoRE alternates between reinforcement learning for the routing policy and contrastive learning for the rule-constrained retriever. The selected rules provide supervision for retriever training, while retrieval performance rewards the routing policy. This closed-loop process enables both components to co-adapt without explicit query-rule supervision. Experiments on STaRK-QA datasets across scientific discovery, biomedical reasoning, and e-commerce knowledge management demonstrate CoRE's effectiveness. Ablations show the benefit of co-adapting the router and retriever. Generation-based routing transfers across datasets and benefits from multi-domain pretraining followed by target-specific fine-tuning. Analysis of rule usage across queries also informs automatic rule refinement and discovery.
Abstract (translated)
URL
https://arxiv.org/abs/2610.14465
PDF
-
Cross-Modal Latent Influence: Patch-Level Training Sample Attribution for Vision-Language Models
2026-10-09 18:04:04Laura Yao, Shixing Yu, Jie Peng, Promit Ghosal, Kyra Gan, Tianlong ChenAbstract
Explaining a vision--language model (VLM) prediction requires tracing training-example influence to the multimodal evidence that mediates it, yet existing attribution methods largely operate only at the sample level. Recent latent-mediated influence methods provide feature-level attribution in language models, but lack the multimodal and spatial structure needed to ground influence in VLMs. We introduce MOSAIC (Multimodal Orthogonalized Sparse Autoencoders for Influence Characterization). It learns a shared sparse image--text latent space whose concepts are 1) sparse within each modality, 2) supported on contiguous image regions, 3) aligned to the phrases that describe them, and 4) approximately decorrelated. It then applies EK-FAC to decompose training-example influence through these concepts and localize their contributions to specific text phrases and image regions. We evaluate MOSAIC on Qwen and Gemma across captioning, chest X-ray understanding, and visual question answering. The learned features are more localized and coherent without degrading task behavior, and ablating influence-ranked latents changes model outputs more than ablating latents ranked by activation or frequency. We find that visual evidence can substantially reorganize training-data attribution while preserving more of the shared latent structure. MOSAIC thus connects where model behavior comes from with what multimodal evidence carries that influence.
Abstract (translated)
URL
https://arxiv.org/abs/2610.14190
PDF
-
Do you 'see' what I 'see'? A Multi-panel Visual Question Answer Dataset for Large Language Model Chart Analysis
2026-10-09 18:00:02Daniel J. Evans, J. P. Naiman, JooYoung Seo, J. Stephen DownieAbstract
Digital libraries increasingly host scientific publications whose figures encode primary research findings, yet making this visual content searchable, accessible, and reusable remains an open challenge. Large language and multimodal models (LLMs/LMMs) provide the possibility to automate many figure parsing tasks such as determining data trends or extracting data from plots. In this work, we consider if these models can reliably support task-oriented visual data analysis when presented with synthetic, but realistic, scientific figures that contain multiple charts. Our Visual Question Answering (VQA) process scaffolds our tasks along a four-level description model from basic conceptual answers to interpretive and analytical understanding. Additionally, we outline our process for synthetically generating our single and multi-panel (charts in which there are multiple panels present) plots and testing the data with various LMMs using quantitative analysis. In general, we find LMMs become less accurate with increasing figure complexity as the number of panels, number of plot elements, and difficulty of questions increase. Finally, because LMM accuracy is expected to change rapidly, our analysis also considers cost relative to performance as a metric to inform future user decisions. We provide the detailed question-answer pairs used in this analysis, along with the code used to generate random plots, question-answer pairs and detailed bounding-box metadata for chart elements. The code and data are available at this https URL and this https URL\_qa\_multipanel
Abstract (translated)
URL
https://arxiv.org/abs/2610.14173
PDF
-
Can Language Models Learn When to Abstain? Policy Learning and Calibrated Answer Selection
2026-10-09 16:33:27Juan Terven, Diana-Margarita C\'ordova-Esparza, Ruben Gomez, Carlos Miranda Rocha, Julio-Alejandro Romero-Gonz\'alez, Edgar Chavez-UrbiolaAbstract
A useful language model should know when to answer and when to say "I don't know." Can we improve that decision by training the model itself or by training a separate component to check its answers? We study both approaches on Qwen3-14B for factual questions, using a score that rewards correct answers, penalizes mistakes, and allows abstention. Across three training seeds and two held-out confidence thresholds, a learned answer selector improves threshold-averaged utility over the original model's calibrated confidence filter. On TriviaQA, mean coverage increases from 20.2% to 30.9%, while the pooled error rate among scored answers decreases from 23.0% to 14.5% under exact-match grading. The tested reinforcement-learning changes do not establish consistent gains. On NQ-Open, the selector improves threshold-averaged utility over the filter but still performs worse than always abstaining. A partial human-reviewed audit preserves the TriviaQA finding. Our results support learning to select answers in this setting while showing that reliable abstention across datasets remains unresolved.
Abstract (translated)
URL
https://arxiv.org/abs/2610.13821
PDF
-
Not All Dimensions Are Equal: Outlier-Aware Attention Linearization with Grouped Quadratic Kernels
2026-10-09 14:40:19Weiqi Sun, Qianyue Wu, Wenjia Wang, Tianyang HuAbstract
Linearizing softmax attention offers a route to efficient Transformer inference, but preserving pretrained attention remains challenging. We identify an important source of approximation error overlooked by standard linearization methods: a small number of outlier dimensions contribute disproportionately to query-key dot products and account for a large fraction of the approximation error. Motivated by this heterogeneity, we propose Outlier-Aware Attention Linearization (OAL), a dimension-aware method for constructing quadratic attention approximations. Starting from a second-order Taylor approximation, we introduce a Hadamard-product formulation that makes individual query-key contributions explicit. This formulation relaxes globally shared coefficients to support dimension-specific linear terms and cross-dimensional quadratic interactions while preserving separable evaluation. To retain this flexibility with fewer parameters, OAL groups dimensions with similar contribution scales to share coefficients, while allowing distinct treatment of the few outliers. The resulting positive quadratic kernel is calibrated against pretrained attention outputs and admits linear-time evaluation with respect to sequence length for fixed head dimensions. On Qwen2.5-0.5B, OAL reduces frozen-attention output relative MSE by approximately 8.9% compared with DiJiang. Replacing middle-layer attention modules and jointly fine-tuning the kernel coefficients with LoRA further lowers WikiText-103 perplexity from 31.71 to 27.50 compared with a learned global quadratic, while improving accuracy on PIQA, HellaSwag, and ARC-E. These results demonstrate the value of outlier-aware dimension grouping for both attention approximation and pretrained-model conversion. The code is available at this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2610.13674
PDF
-
Inject a Turn, Not a Token: Learning Where to Redirect LLM Reasoning for Self-Correction
2026-10-09 14:37:51Jiwon Jeon, Myungsik Cho, Jeonghye Kim, Youngchul SungAbstract
Large language models struggle to reliably identify and correct errors in their own reasoning. Unlike prior training-based approaches that fine-tune an LLM to generate revisions or feedback, we learn where to intervene in failed rollouts and determine how to redirect them. To choose where to intervene, we train a 130K-parameter MLP head on a frozen encoder using only final-answer correctness. For redirection, we find that under a fixed generation budget, the tested inline injections fail to reliably redirect reasoning, whereas presenting the same instruction in a new user turn changes the model's course. The resulting correction framework, UPTURN (User-Prompt TURN injection at a learned position), requires no LLM fine-tuning: it retains a selected prefix, injects an instruction as a new user turn, and regenerates the suffix. Across four reasoners and six math benchmarks, UPTURN improves average accuracy in every setting over a 32-rollout baseline. It outperforms self-correction and reflection-token injection baselines under matched correction budgets, transfers across reasoners and to GPQA-Diamond without retraining, and complements GRPO and self-correction methods.
Abstract (translated)
URL
https://arxiv.org/abs/2610.13672
PDF
-
Benchmarking Answer-Grounding Consistency in Remote Sensing Visual Question Answering
2026-10-09 14:15:50Ping Guo, Zhiqi Huang, Xinran Li, Linkun Yu, Jingyin Pei, Kamata SeiichiroAbstract
In grounded remote sensing visual question answering, a model produces a textual answer and identifies relevant objects in the image as grounding evidence. Existing evaluations measure answer correctness and localization quality, but do not directly assess answer-grounding consistency. We propose the Answer-Evidence Support Protocol (AESP) to assess this consistency through evidence support: whether the grounding evidence supports the model's own answer. AESP applies task-specific rules to existence, counting, and comparison questions. We further construct the Grounded Remote Sensing Visual Question Answering (G-RSVQA) benchmark, containing 15,027 samples with questions, reference answers, and all related object-level annotations. Experiments on seven grounded vision--language models reveal failures of evidence support across all three question types. Up to 50.82% of correct answers in the evidence evaluation subset are not supported by the provided grounding evidence.
Abstract (translated)
URL
https://arxiv.org/abs/2610.13636
PDF
-
AgroGround: Multi-Granularity Grounded Recognition in Agriculture
2026-10-09 11:13:59Abdulla Alshehhi, Zongyan Han, Rao Muhammad AnwerAbstract
Agricultural visual models are typically evaluated for either recognition or localization, but reliable diagnosis requires identifying what is present and localizing the evidence. Agricultural visual question answering (VQA) datasets carry rich semantic labels but rarely link them to image regions, and adding such annotations by hand is costly at scale. We introduce AgroGround, a large-scale dataset for grounded agricultural recognition: identifying plant diseases and other agricultural targets and localizing their image regions. An automated pipeline converts the labels of eight agricultural VQA datasets into annotations for disease lesions and whole objects, producing 794,850 instruction examples. Healthy images provide negative supervision for disease queries, teaching the model to return empty predictions. We fine-tune a shared vision-language model on known-target grounding instructions combined with instructions requiring both recognition and localization. We evaluate predicted identities, regions, joint correctness, and healthy-image abstention on 1,480 human-verified images disjoint from all training data. Grounding-only fine-tuning reduces recognition accuracy from 51.8\% to 29.1\%, while adding recognition-and-localization instructions raises it to 72.6\%. With images and annotations held fixed, combining the two formats raises joint accuracy from 19.2\% to 43.3\% at comparable grounding. Healthy negatives raise abstention on healthy images to 95.0\%, and reinforcement learning improves lesion-level grounding. The resulting 2B model exceeds its annotation teacher in grounding F1 on our benchmark and on the external PlantSeg test set. AgroGround establishes a benchmark for grounded agricultural recognition, measuring joint correctness of identity and localization along with abstention on healthy images. The code is available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.04425
PDF
-
PRIME: Pre-Retrieval Index Modeling Enhancement for Structured Document Retrieval-Real
2026-10-09 09:44:47Cheuk Hei Chung, Li LiuAbstract
High-precision retrieval over external knowledge bases is essential for LLM agents to operate effectively. However, existing Information Retrieval (IR) research primarily optimizes query-side retrievers (e.g., ColBERT, BGE) and generators (e.g., Self-RAG, GraphRAG), while reducing corpus-side representation to chunking. This leaves index-side quality an underexplored lever. We propose Pre-Retrieval Index Modeling & Enhancement (PRIME), an index-side optimization framework built on two complementary tracks: an LLM-driven Semantic Restructuring (SR) track that transforms long documents into structured, information-dense sections, and a deterministic Metadata Synthesis (MS) and Hierarchical Retrieval (HR) track that extracts semantic anchors and heading structure for coarse-to-fine section retrieval. Our evaluations show that deterministic metadata matches structure-aware chunking's retrieval quality while cutting embedding computation by 45.8 $\times$ and index vector count by 22.4 $\times$. On top of these gains, Semantic Restructuring (SR) further improves Answer-F1 by 13.7% in full deployment, with characterized quality-fidelity-cost trade-offs. Across QASPER, BRIGHT, and NFCorpus, we further characterize three qualitative regimes that determine whether index-side optimization helps or degrades retrieval, while showing that these index-side effects are neutralized under a strong reranker.
Abstract (translated)
URL
https://arxiv.org/abs/2610.13296
PDF
-
Do Prompt Optimizers Break Requirements They Are Not Scored On? A Preregistered Study
Abstract
Prompt optimizers rewrite a prompt to raise one target metric, but prompts also state requirements the metric does not score, such as an output format, citations, or abstention. In a preregistered study, we run GEPA, MIPROv2, and TextGrad on HotpotQA, SQuAD 2.0, and IFEval and count per-item violations of unscored requirements with deterministic checks. By the preregistered criterion, all three optimizers produce regressions, substantially on HotpotQA, where net regression in citation validity reaches 10.7 to 41.6 points while exact match rises by 21 to 34 points. A post-hoc decomposition changes this reading. The optimized prompts elicit reasoning before the required JSON object, and under the 200-token output cap used throughout, 13.9% to 47.0% of responses end before any answer appears. Among parseable outputs, net regression is within 3 points of zero. With a 1,024-token cap, extraction failures disappear, content regression stays below 1 point, and exact match rises further. Two simple gates fail the preregistered criterion. The side effect is narrower than hypothesized: optimized prompts are more accurate and break an explicit output contract, and a lenient target metric does not show the cost.
Abstract (translated)
URL
https://arxiv.org/abs/2610.13287
PDF
-
Complexity-Calibrated Reconstruction Discrepancy for No-Reference Point Cloud Quality Assessment
2026-10-09 08:35:28Haocheng Wu, Xiangjie Sui, Xin SunAbstract
The discrepancy induced by surface reconstruction in a distorted point cloud is jointly influenced by its intrinsic structural complexity and signal degradation. This observation motivates a referencefree point cloud quality assessment (PCQA) strategy that estimates and compensates for the complexity-induced component. We implement this strategy through successive reconstruction. Specifically, reconstructing the distorted point cloud produces an input-toreconstruction discrepancy that combines the effects of signal degradation and content complexity. Reapplying the reconstruction operator with stronger smoothing produces a discrepancy between successive reconstructions, which serves as a proxy for the complexityinduced component. This proxy is then used to calibrate the inputto-reconstruction discrepancy. We apply the same strategy to 3D geometry patches and their aligned 2D regions, enabling local comparison of the resulting discrepancies. The calibrated geometric and projected-view discrepancies are then fused into a final degradation score. For both branches, the discrepancies are measured using frozen pretrained perceptual representations. Our method does not require PCQA-specific training or fine-tuning. Experiments on three public databases show that the proposed method consistently outperforms most generic baselines and achieves performance comparable to several full-reference this http URL code is available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.13219
PDF
-
PreCoG: Preview-Edit for Efficient Video Correction in Generation-Time
2026-10-09 08:01:49Yilai Liu, Wenqi Pei, Zhifeng Lu, Shiyuan Zhang, Zongxi Liu, Hongyang DuAbstract
Iterative text-to-video generation relies on feedback to correct compositional and temporal failures. However, existing methods typically require another full generation run for each correction, repeating computation and potentially disrupting already correct content. We introduce PreCoG, an iterative framework for correcting semantic failures during video generation. PreQA diagnoses failures from low-cost previews and constructs corrective conditions, while PreEdit edits cached generation trajectories under these conditions, attenuating orthogonal changes to limit disruption of already correct content. PreCoG completes generation only for the selected candidate, reusing its cached trajectory to reduce redundant computation and help preserve semantic consistency between the preview and the final video. Experiments show that PreCoG achieves the highest VBench and T2V-CompBench scores and approaches VSR's leading VS2 score at $0.56\times$ its inference time. It also outperforms GENMAC on all three quality and alignment metrics at $0.86\times$ its inference time.
Abstract (translated)
URL
https://arxiv.org/abs/2610.13174
PDF
-
Locate, Judge, Read: Role-Specialized Tool Agents over a Knowledge-Point Index for Open-Domain Multimodal Document QA
2026-10-09 07:21:36Wenlin Lin, Fanzhe WeiAbstract
Question answering over large heterogeneous document collections requires first finding the right passages and then reading them, yet current retrieval-augmented agents give one model both kinds of tools and let it schedule itself. We present KnowForge, a pipeline separated by job: a hybrid-retrieval knowledge-point index; a Locator agent that can only search; a code-scheduled bridge judge that forces the missing second hop; a typed, size-bounded evidence pool that merges every search's ranking; a Deep Reader agent that can only read text, images and tables; and a synthesizer with a format checker. With a self-hosted Qwen3.6-35B-A3B (3B active), KnowForge reaches F1 47.5 on the full open-domain M3DocVQA (2,441 questions, 3,368 PDFs), +11.0 over the published M3DocRAG system and +7.1 over one-shot retrieve-and-answer on the same index and model, and 53.5 with DeepSeek-V4-Flash. On MultiHop-RAG it reaches 73.5 accuracy, above the strongest published RAG and GraphRAG results (71.2); on the single-document MMLongBench-Doc it lifts the same model from 34.8 to 47.0, level with GPT-4o reading page images (46.3). Per-job ablations on the full benchmarks attribute the gain: the Deep Reader is worth 3.8, 9.8 and 0 points on the three corpora, tracking how visual each is; merging both agents into one with all tools and twice the rounds loses 1.9 F1 because it keeps searching instead of reading; a round-budget sweep shows a step at four Locator rounds and a plateau beyond; query latency scales with model calls rather than corpus size. We serve the three indexed benchmarks as read-only knowledge bases (free registration) and release all prompts and the runner code: this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2610.13128
PDF
-
BenCoDiC: A Large-Scale Contextual Cyberbullying Corpus for Bengali Political News Discourse FULL VERSION
2026-10-09 07:16:29Md. Nesarul Hoque, Rudra Pratap Deb Nath, Umme Salma, Shima Chakraborty, Md. Hanif SeddiquiAbstract
Cyberbullying is a growing concern in online communities, yet existing Bengali corpora remain limited in contextual information, annotation transparency, scale, accessibility, and support for multidimensional analysis, thereby hindering robust detection and comprehensive study of harasser behaviors and networks. To address these gaps, we introduce the Bengali Contextual and Diverse Cyberbullying (\textit{BenCoDiC}) corpus, a publicly available contextual cyberbullying resource for Bengali political news discourse comprising 50,250 textual instances collected from 100 YouTube videos and their associated discussions. The corpus preserves caption--comment--reply relationships and is organized through a nine-dimensional star schema covering users, videos, textual content, context, time, engagement, political parties, target identities, and language categories. Five AI chatbots were initially evaluated for annotation but did not meet the predefined reliability criterion. Consequently, a two-tier human annotation process with expert verification was adopted, achieving Fleiss' $\kappa=0.86$ and Krippendorff's $\alpha=0.86$. Statistical analyses characterize the corpus, while transformer-based experiments show that incorporating conversational context yields modest but consistent improvements over isolated-text classification. \textit{BenCoDiC} supports context-aware cyberbullying detection and perpetrator-oriented behavioral and network analysis. The corpus is available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.13121
PDF
-
LongHarness: Advancing Agentic Reasoning over 100M-Token Corpora via Stateful Evidence Management
2026-10-09 06:32:02Songtao Huang, Zhijie Zhong, Xiaohan He, Shuaiyu Zhang, Zhiyin Yu, Li Kang, Heng Zhou, Yutao Fan, Xiangchao Yan, Bo Zhang, Shiyang FengAbstract
Corpus-level long-context reasoning requires agents to acquire and integrate evidence across document collections that exceed their context windows. Existing agent harnesses enable incremental access to these corpora, but acquired evidence and evolving conclusions often remain entangled in a growing interaction history. This implicit state management can lead to two coupled failure modes: unbounded exploration, where retrieval expands the interaction history without resolving decision-relevant information gaps, and evidence-state drift, where previously acquired evidence or subsequent corrections fail to inform the current conclusion. We present LongHarness, a model-agnostic harness that makes the interaction between acquired evidence and evolving conclusions explicit. Its core mechanism, Evidence-Conclusion Coevolution, uses unresolved conclusions to guide exploration toward missing information, while newly acquired evidence strengthens or revises the conclusions that directed the search. LongHarness couples this process with Evidence Lifecycle Management, which keeps accumulated evidence current and traceable throughout the interaction. Together, these mechanisms focus retrieval on remaining information needs and keep conclusions aligned with current evidence. We evaluate LongHarness on OfficeQA Pro V1 and V2, AA-LCR, and LongBench v2 using GLM-5.3-Flash and Qwen3.8-Flash. Compared with direct reasoning using the same backbone, LongHarness improves GLM-5.3-Flash exact match on OfficeQA Pro from 15.8% to 60.2% on V1 and from 4.4% to 61.1% on V2. Against state-of-the-art agent harnesses, it surpasses Claude Code by 12.0 percentage points on V1 and Codex by 18.9 points on V2. With Qwen3.8-Flash, LongHarness reaches 61.7% on V1, outperforming Claude Code by 8.3 points and Codex by 12.0 points.
Abstract (translated)
URL
https://arxiv.org/abs/2610.13075
PDF
