Papers

Can language agents continually learn from experience, turning earlier interactions into reusable capabilities? AhaBench evaluates this ability through exploration after solved hidden-state puzzles, computational transfer after mathematical teaching, and sustained business operation under delayed feedback. The benchmark is agnostic to how an agent learns; the evaluated agents use fixed model weights. Curriculum profiles, teaching contrasts, and daily trajectories reveal a common challenge: using explicit guidance is more reliable than generalizing beyond it or sustaining useful behavior. Across the Puzzle panel, the advantage over matched cold targets is 36.0-53.5 points greater with trace support than at the trace-free endpoint; Qwen 3.6 Plus nevertheless retains a +12.57-point post-curriculum gain. In Euler, worked procedures yield 80.0-100.0% held-out accuracy across models, while question-plus-answer teaching yields 0.0-73.9%. Vending trajectories separate sustained profit, late recovery, and incomplete operation: Doubao Seed 2.0 Pro finishes nominal operation at +495 but averages -10 over the year. Together, these results make continual learning an operational target: experience should yield capabilities that remain effective as guidance, inputs, and business states change. We release tasks, validators, a simulator, records, and analyses for developing agents that turn useful insights into lasting abilities.

Large Language Models are increasingly deployed for sophisticated data engineering tasks such as generating structured queries from natural language, Text-to-SQL, and automating complex spreadsheet operations. However, maximizing their utility demands both higher finetuning-free accuracy and solutions to the computational bottleneck imposed by the Transformer architectures inherent quadratic (On2) time complexity. This paper introduces a novel drop-in neurosymbolic layer designed to seamlessly integrate into existing LLM backbones enhancing logical reasoning and mitigating long-context resource consumption. On the reasoning front, the layer immediately and significantly improves performance yielding an average accuracy increase of 85% across rigorous benchmarks including BIRD-CRITIC and LiveSQLBench, critically achieving these gains without any task specific finetuning or RLHF. Concurrently, we repurpose this approach to address the severe computational strain of long context inference. By leveraging symbolic processing to prioritize and compress relevant contextual information the layer reduces the effective token usage by over 50% and brings the effective time complexity down from O(n2) to approximately O(n) on certain long context tasks. This dual impact approach not only makes LLMs substantially more reliable for data engineering but also drastically reduces the computational pressure on inference chips, making long context tasks more manageable and cost effective.

LLM-powered meeting assistants are deployed at scale, yet systematic evaluation of their grounding fidelity remains limited to static benchmarks that miss failure modes tied to specific discourse structures or reasoning demands. We propose Evaluation-as-Search (EaS), a feedback-driven methodology that frames quality evaluation as an adaptive search over the space of natural questions a meeting participant might ask. Rather than sampling uniformly, EaS learns from evaluator feedback across iterations to concentrate probing effort on cognitive demands where failures are most likely, guided by a UCB-scored coverage map and blind multi-dimensional quality evaluation. Using EaS, we construct MeetingProbe, a benchmark of over $3{,}000$ annotated question–answer pairs spanning 20 transcripts from three meeting genres and three LLM assistants. In ablations, adaptive search surfaces $2.5\times$ more failures than random probing ($7.1\%$ vs. $2.9\%$ finding rate), with the strategic planner contributing the largest individual effect. Across three models, we observe a clear capability gradient and identify eight recurring failure categories dominated by discourse-pragmatic challenges rather than factual recall errors. We further validate MeetingProbe across multiple model families and providers, finding a clean capability gradient and a curated subset of universal failures that no model handles. MeetingProbe is released publicly to support reproducible evaluation of meeting assistant grounding fidelity.

Most legal NLP resources draw from federal case law and focus on coarse classification, leaving administrative adjudication, where the vast majority of government decisions occur, essentially unaddressed. We introduce ImmigrationReason, a large-scale structured dataset derived from 12,375 non-precedent decisions of the U.S. Citizenship and Immigration Services (USCIS) Administrative Appeals Office (AAO) spanning 2005 to 2026. Each record captures the applicable legal framework, per-criterion evidence-sufficiency findings under a five-category label, verbatim adjudicator-criticism quotes, all citations, and final dispositions, alongside high-quality Claude-transcribed source text. Extraction quality is validated through a three-pass pipeline combining two independent modalities with comparison-prompt adjudication by Opus 4.7, and verified by domain experts on a 500-record sample. The dataset documents nearly 9,000 verbatim instances of AAO-identified legal errors, spans a natural legal-regime transition (the 2016 Dhanasar rule change), and covers 21 years of adjudication. We analyze the dataset in detail and outline research directions it enables, from outcome prediction and adjudicator-error analysis to agent design for high-stakes regulatory domains.

General-purpose large language models (LLMs) are increasingly used to answer religious questions, but for Islamic content they carry two serious risks: factual fabrication (inventing Qur'anic verses or hadith) and subtle value misalignment. We present Ansari, a deployed, retrieval-grounded Islamic AI assistant that has handled more than 140,000 conversations across 25+ languages since June 2023. Ansari is built around an agentic retrieval loop: a tool-using language model issues searches against authenticated Islamic corpora – the Qur'an, hadith collections, a multi-volume jurisprudence (fiqh) encyclopedia, and exegetical (tafsir) sources – and answers only on the basis of what it retrieves, with citations attached for verification. We describe the system's architecture (the agent loop, the retrieval tools, the corpora, and the system prompt that encodes editorial and theological policy), its multi-platform deployment (web, mobile, WhatsApp, and as a Model Context Protocol server and an Agent Skill), and what 140,000 real conversations reveal about how Muslims actually use such a tool. We report results on several complementary evaluations – zero-shot performance on accredited institutional exams, a human-rated validation during Ramadan, and two independent, externally run benchmarks on which Ansari currently tops the public IslamicMMLU leaderboard ahead of frontier models and is competitive on Islamic legal reasoning (IslamicLegalBench) while strongly resisting false premises – and draw out lessons that generalize beyond Islam to any faith- or values-sensitive deployment of LLMs: grounding is necessary but not sufficient, the system prompt is a theological as much as a technical artifact, and the absence of community in how models are formed remains a hard gap.

Microservice placement in the compute continuum is driven by low-level Service-level Objectives (SLOs), but requiring users to specify metric-level constraints creates an adoption barrier and increases misconfiguration risk. Although large language models (LLMs) can interpret natural-language intents, direct generation of orchestration-consumable SLO artifacts remains unreliable due to unsupported constraints, incorrect grounded values, and schema violations. These errors can propagate to downstream placement logic and produce infeasible or incorrect placements. This paper presents Intent Engine, a natural-language intent translation architecture that constructs validated SLO artifacts for compute-continuum service placement. Intent Engine acts as an intent acquisition and SLO construction layer for existing intent-driven orchestration and placement frameworks; it does not perform placement or runtime QoS optimization. The architecture combines schema-constrained extraction, retrieval-grounded value construction from monitored infrastructure state, and validation against supported constraints before emitting the final SLO artifact. We evaluate Intent Engine using a 716-record intent-to-SLO dataset derived from an edge-cloud testbed, including valid and invalid intents. Across GPT-4.1 mini, Claude Sonnet 4.5, and DeepSeek V4-Flash, Intent Engine outperforms prompting baselines and a non-LLM rule-based parser. With GPT-4.1 mini, it achieves 0.941 total F1 Score and reduces aggregate hallucination by 85.1%, while lowering downstream placement failure from 30.8% to 2.1%.

While recent text-to-speech (TTS) models achieve high naturalness, controlling fine-grained expression via natural-language instructions remains challenging. We introduce Poly- InstructTTS, which learns expressive speech from open-ended instructions using in-the-wild audiovisual data. We build a scalable multi-modal pipeline to construct a 1,000-hour instruction-annotated corpus covering 1,000+ fine-grained emotions and styles. The framework uses a prompt-free GPT with attribute-based thinking tokens, followed by a flow-matching module that injects timbre from a reference audio. We also present a speaker fine-tuning procedure to transfer instruction control to specific speakers while preserving persona. We further extend InstructTTSEval with broader tasks. Experiments show that Poly-InstructTTS delivers strong performance in instruction adherence and expressiveness. Audio demos and the expanded testset are available on our project page.

Systematic reviews rely on quality appraisal of included studies, a process that is time-consuming and sensitive to ambiguity in checklist criteria. Although large language models (LLMs) offer opportunities to support these tasks, appraisal checklists are typically treated as fixed inputs, and it remains unclear how their design affects agreement with expert judgments. Therefore, we investigate (1) whether LLMs can approximate human judgments in checklist-based appraisal and (2) whether patterns of human-LLM disagreement can be used to identify and improve ambiguous checklist items. Using the Guidelines for Reporting on Latent Trajectory Studies (GRoLTS) checklist, we compare LLM-generated assessments with expert annotations across three research topics and two checklist versions. Agreement is assessed using item-level accuracy, chance-corrected agreement, and preservation of study-level rank ordering. We find that performance varies substantially across checklist items, with ambiguous and conditional criteria producing the greatest disagreement. Revising these items improves both raw and chance-corrected agreement. Although item-level misclassifications persist, LLM-generated scores often preserve the relative ranking of studies when high-agreement items are retained. These results indicate that reliable LLM-assisted appraisal depends not only on model choice but also on checklist design. The findings suggest that analyzing human-LLM disagreement can help identify problematic checklist items and support the iterative improvement of research synthesis workflows.

Token-level hallucination detectors score each token independently from a single signal, and fail exactly when the generating model is confidently wrong. This paper instead treats hallucination as a temporally extended span and detects it by sequence labeling: each token is scored from a 33-dimensional feature stream that fuses text statistics, Natural Language Inference (NLI) entailment, and language model surprisal, with no access to model internals. A Bidirectional Gated Recurrent Unit (BiGRU) over these features reaches an AUC of 0.840 on RAGTruth (10 seeds), an 11-point gain over an independent logistic-regression baseline (p = 0.002, Wilcoxon signed-rank). A controlled decomposition attributes most of the gain to temporal order rather than model capacity: evidence propagates from confident positions to ambiguous neighbors within a span. The same 0.845 ceiling recurs across recurrent, state-space (Mamba), and attention architectures, locating the bottleneck in the feature set rather than the model. Because it reads only the generated text and external signals, the detector works on closed-source models, and it keeps working on text produced by language models it never saw during training, losing under 4% AUC.

On-policy self-distillation (OPSD) teaches large language models new skills through a teacher that shares the student's backbone and supervises its own rollouts. Existing teachers either inject privileged context at the input – inducing post-hoc rationalization – or fine-tune weights, accumulating drift and forgetting across tasks. We propose \method, whose teacher differs from the student only by a learnable soft prompt: trained on $(x, y_\text{gold})$ pairs with the backbone frozen, the prompt yields a task-specific teacher that preserves the student's exact representational geometry. \method\ extends naturally to multi-task settings by routing each example in a merged corpus to its corresponding soft-prompt teacher, allowing a single student to absorb knowledge from $K$ teachers in parallel; at inference, all prompts are discarded. On Qwen3-1.7B-Base and Phi-4-mini-instruct across four tasks (Science, Tool Use, Biology, Math), the single-task variant (OPD with a PT teacher) matches or exceeds full fine-tuning while training orders of magnitude fewer parameters, and the multi-task variant achieves the best overall average ($56.2$ on Qwen3-1.7B-Base) while preserving general-capability benchmarks – in contrast to sequential SFT, which degrades both.

Bigger language models are less reliable. Across three families, three benchmarks and six rungs, including in-the-wild chat logs, scaling closes the start-of-response knowledge gap up to $7\times$ while within-response knowledge degradation grows up to $39\times$. We trace that residual to one variable, the per-position disagreement $\delta = \log p_M - \log p_O$ against a stronger oracle, whose second moment splits exactly into bias$^2$ $\mathrm{KL}(p_M \,\|\, p_O)^2$ and decoding risk $\mathrm{Var}[\delta]$. That split is an interpretability statement before it is a statistical one: the model's self-readable uncertainty $H(p_M)$ enters only the bias term, so the risk term has no model-readable component. Risk also takes a growing share of the squared error with scale, $31\%$ to $49\%$ from $1.7$B to $14$B. At a fabrication $H(p_M)$ relaxes within one token while risk persists up to $23\times$ longer, leaving a confident-but-precarious regime that bridges consecutive fabrications ($+69\%$ at $14$B). Contracting that risk at fixed $\mathrm{KL}$ removes $35$-$74\%$ of web-verified hallucinations across six rungs and three families. Semantic entropy fires $\approx$$30\%$ less on that branch ($p\!<\!10^{-16}$) though it carries nearly $4\times$ the fabrications. Bigger models snowball mistakes faster, through a failure mode that is dominant, self-perpetuating, causal and invisible to the model itself.

The transmission chains (sanad) of Islamic Hadith literature encode relationships among tens of thousands of historical narrators whose biographical records are dispersed across independently maintained digital databases that share no common identifier. We present a two-phase entity resolution pipeline that links narrator names from the Sanadset 650K corpus - 650,986 Hadith records from 926 books containing 185,216 unique narrator name variants - to two biographical databases: Hadithtransmitters (Hawramani; 100,915 entries) and Muslimscholars (25,247 entries). Phase 1 matches Sanadset names to Hawramani using name-only similarity (Sanadset carries no metadata), yielding 94,628 links (51.1%; HIGH 39,938 / MED 54,690). Phase 2 cross-references Hawramani against Muslimscholars via a weighted multi-signal function combining name similarity, death-year proximity, and reliability grade polarity, yielding 95,573 links (94.7% of Hawramani; HIGH 18,245 / MED 71,546 / LOW 5,782). Chaining the two phases gives Sanadset narrators transitive access to Muslimscholars data. The linked data enable construction of a 185,216-node, 814,093-edge directed transmission graph enriched with cross-source biographical metadata. The annotated link corpora and enriched graph are released as open resources.

Current large-language-model (LLM) physics benchmarks are usually scored by answer accuracy, which cannot distinguish genuine reasoning from recall of familiar problem patterns and reveals little about where a model's reasoning breaks down. We introduce an auditable four-stage diagnostic that evaluates whether an LLM can reason inside an unfamiliar physics framework through induction, formulation, prediction, and review. The diagnostic combines locked pre-registrations, fresh sessions between stages, dual-LLM judging, and a human-audit pathway, and we apply it to three parallel physics worlds: a single-equation counterfactual world ($F=mv$), a historical framework (Aristotelian mechanics), and a four-domain counterfactual world (Decay World). Across Claude Opus 4.7, GPT-5.5, and Gemini 3.1 Pro, the three worlds yield composite PASS rates are 6/15, 6/15, and 0/15 respectively (content $\land$ structural for $F=mv$ and Aristotelian, content axis only for Decay World where the structural axis is out of scope). The most pointed empirical pattern is a qualitative-versus-quantitative asymmetry: in Decay World, models almost never predict the wrong direction of change, but frequently compute the wrong ratio by slipping back to standard-physics relations. The protocol also surfaces two methodology findings: LLM-judge reliability does not transfer across frameworks, and Stage 4 self-review is weak in every framework, with the model's own review wrongly reporting no earlier error in at least two-thirds of the trials that actually contained one. We release the full prompts, responses, verdicts, and audit records.

Effective writing feedback is among the strongest drivers of student learning, yet producing it at scale is labor-intensive. LLMs offer a natural path to scaling writing support, but two gaps stand in the way: few public corpora capture how instructors actually deliver feedback in real classrooms, and no reliable method measures whether generated feedback aligns with what an instructor would write. We address both. SEFORA is a public corpus pairing instructor inline feedback with assignment prompts, rubrics, scores, and multi-draft revisions across various college writing genres, comprising 564 drafts and 8,240 instructor annotations. UniMatch is a reference-based evaluation framework for open-ended generation: it segments feedback into feedback units, scores their semantic correspondence under instructor-derived criteria, and aligns them via optimal matching to yield interpretable precision, recall, and F1. Across 74 experimental configurations spanning multiple LLMs, no setting exceeds 0.4 F1. UniMatch reveals that models struggle to identify the feedback instructors would prioritize, and performance degrades as models generate more.

Maltese has substantial text corpora and pretrained language models, but paragraph-scale OCR training data remains scarce; NOMOCRAT provides 57 verified annotated pages. LV-ROVER-MLT combines synthetic fine-tuning of Tesseract~5 with five complementary recognition streams and lexicon-gated word-level arbitration adapted to Maltese diacritics and hyphenation. In the DocEng~2026 Maltese OCR competition, the system placed first with held-out CER 0.0074; the next-ranked submission scored 0.0161 and NOMOCRAT scored 0.0163. The same approach produced a significant improvement over stock Tesseract on Luxembourgish, while the Hungarian result was inconclusive. A 36,803-pair Maltese OCR corpus constructed from EUR-Lex and Wikipedia provides an additional paragraph-level resource. Code, model weights, and corpus data are public.

How do two agents invent a shared language from scratch? In a Lewis signaling game, a sender and receiver must coordinate on a code using only their interaction history. We study five memory architectures across varying channel configurations with LLM agents and find that memory architecture matters more than channel capacity. Agents with a persistent private notebook benefit from surplus channel capacity and avoid the high-capacity collapse seen in stateless agents, achieving the most reliable coordination ($0.867 \pm 0.023$ at capacity = 25). Stateless agents peak at moderate capacity and then degrade as the vocabulary grows beyond what a rolling context window can track The notebook externalizes learned conventions, freeing agents from having to re-derive codes each round. An information bottleneck-inspired argument predicts an optimal capacity equal to the number of objects. Instead, the bottleneck (capacity = 8) proves to be a fragility point, and surplus capacity is generally better. We show that channel capacity alone cannot predict coordination; memory architecture determines whether agents turn interaction history into stable conventions, and both dimensions are needed to understand how signals become language.

Reinforcement learning for diffusion large language models (dLLMs) has largely moved to trajectory-aware methods. The current state of the art, TraceRL, holds that random masking is mismatched with the model's inference trajectory, and it reconstructs that trajectory during training by slicing each rollout into up to K/s trajectory-aligned training samples, a cost that grows with the block size K. We show that this mismatch can be mitigated without reconstructing the trajectory. Our method, SLIM-RL, bounds the commit risk of each rollout step with a tau-budget decoder, reducing aggregate commit risk in the training data. During optimization, SLIM-RL trains on these risk-controlled rollouts with a trace-free random-masking objective that adapts variance-reduction tools, combining sequence-level importance sampling, deterministic quadrature over masking levels under a mean-preserving, monotonically decreasing per-block mask schedule that we introduce. On SDAR-4B, SLIM-RL matches TraceRL's best MATH500 accuracy on only 0.46x its training samples at block size 16, improving over TraceRL by 6.32% on MATH500 and 11.05% on GSM8K under matched dynamic sampling. At block size 4, the 4B SLIM-RL surpasses the larger LLaDA-8B and Dream-7B dLLMs on math, exceeding LLaDA-8B by 10.76% on MATH500 while staying below the autoregressive Qwen2.5-7B. On code, it improves over TraceRL by 4.20% on MBPP and 3.65% on HumanEval. The tau-budget decoder transfers training-free across LLaDA, Dream, and SDAR. The source code is available at https://github.com/laolaorkkkkk/SLIM-RL .

Khipus – knotted cord devices – were the primary recording medium of the Inka Empire (c. 1400-1532 CE), yet their system remains undeciphered. We present a reproducible machine-learning pipeline applied to the Open Khipu Repository (OKR), a public database of 619 khipus comprising 54,403 cords and 110,677 knots. We engineer 27 structural features per khipu and apply (i) unsupervised clustering via UMAP and HDBSCAN, recovering three structurally distinct groups (silhouette = 0.769); (ii) supervised provenance classification via gradient boosting, reaching F1 = 0.86 for the Inka Late Horizon imperial style; and (iii) SHAP-based interpretability, which identifies cord twist direction as the dominant structural discriminator of imperial khipus. We further report two findings of methodological interest. First, one cluster is dominated not by a geographic region but by nineteenth-century European museum collections, indicating that colonial acquisition and recording practices are structurally encoded in the corpus. Second, we provide an independent computational verification of the recto/verso (moiety) structure of the six Santa Valley khipus reported by Medrano and Urton (2018), reproducing both the aggregate attachment ratio and the identification of the single mixed specimen–using only the public OKR database, without physical access to the objects. We additionally report a negative result: knot-type sequence order, encoded as n-grams, adds no provenance signal beyond aggregate features. All code and data are openly available.

Text embeddings are standard for semantic similarity tasks, yet their evaluation remains an open challenge. Current benchmarks are static, cover only a limited set of languages, are often domain-specific, susceptible to overfitting, and poorly representative of low-resource languages. To address these limitations, we introduce ALEE, a framework that extends Sentence Smith (Li et al., 2025) to the cross-lingual and paragraph level. ALEE uses Abstract Meaning Representations (AMR) to generate English minimal pairs with controlled, fine-grained semantic shifts, which are paired with translations in target languages. This approach enables targeted diagnostics for models in any language with English parallel data. We conduct a large-scale empirical study across a diverse set of embedding models and 275+ languages spanning three parallel datasets. On ALEE, performance varies substantially across languages, text lengths, and linguistic phenomena, exposing persistent gaps in cross-lingual semantic representation that track language prevalence in training resources and subword tokenization. We release ALEE at https://github.com/Andrian0s/any-lang-embed-eval

Knowledge-Based Visual Question Answering (KB-VQA) aims to evaluate whether Visual Language Models (VLMs) can retrieve, ground, and reason over external structured knowledge beyond visual evidence. In practice, answer accuracy is widely adopted as the primary evaluation metric, implicitly treating correctness as a proxy for knowledge-grounded reasoning. However, for existing KB-VQA benchmarks, this proxy relies on critical assumptions that are often overlooked and rendered unreliable by benchmark issues: annotated answer must be derivable from the associated knowledge base, question must be well-posed with sufficient constraints, and visual setting must meaningfully require grounded disambiguation. In this work, we show that these assumptions are systematically violated in existing KB-VQA benchmarks. Our audit reveals substantial instances with missing or contradicted answers and underspecified questions that render accuracy a misleading metric. Furthermore, we find that existing datasets rely on visually trivial, single-entity scenes that bypass the need for sophisticated visual-to-knowledge mapping. We demonstrate that even with controlled architectures, these flaws lead to distorted model rankings and overestimations of reasoning capabilities. To address this, we introduce (1) a principled audit-and-repair protocol that restores answer derivability and question clarity, and (2) a controlled multi-entity augmentation protocol that introduces visual ambiguity to challenge initial retrieval and grounded reasoning. Re-evaluation under corrected and augmented settings yields markedly different performance trends. Our findings call for rethinking evaluation protocols and designing more interaction-aware KB-VQA benchmarks that prioritize verifiable reasoning over simple matching.

Hallucination remains one of the central obstacles to deploying medical LLMs. Yet, even when hallucination can be detected, it is still unclear whether the internal representations associated with it can be used for control rather than detection alone. Using four open-source models across a suite of medical question-answering datasets, we show that a simple, carefully conditioned probe can reliably detect hallucination, with AUROC scores between 0.77 and 0.86 in our case. We further show that this signal is distributed and redundant rather than narrowly localized. Systematically selected neurons outperform random neurons only at very small subset sizes, whereas random subsets of a few hundred neurons recover nearly the full signal, and low-dimensional random projections preserve most of the detection performance. Beyond detection, we test whether this representation is causally actionable. Across 16 model–dataset combinations, our results reveal a sharp gap between decodability and controllability. The same internal structure that makes hallucination easy to detect does not translate into reliable neuron-level control. These findings show that medical hallucination seems to be readily visible in internal activations, but not easily corrected by steering the neurons most associated with it. More broadly, our results suggest that hallucination mitigation is not simply a matter of identifying the right neurons, and point to a deeper separation between what representations reveal and what they allow us to change.

Three of the most popular methods for training language models to reason look like three different tricks. They are not. All three adjust a single number: standard deviation, reflecting how much a prompt's sampled answers disagree. When such a model is trained, it answers each problem many times, and an automatic checker marks every answer right or wrong. The standard deviation of those marks measures the disagreement: largest when the answers split evenly between right and wrong, and zero when they all agree. Group Relative Policy Optimization (GRPO) divides by this number, GRPO Done Right (Dr. GRPO) drops the division, and Decoupled Clip and Dynamic Sampling Policy Optimization (DAPO) discards the groups where it is zero. Each is presented as its own fix, yet this paper proves they are three settings of one dial. That dial is not cosmetic: for right-or-wrong rewards, the disagreement is exactly the size of the training update, the group-standard-deviation identity. A split group teaches the most, while a unanimous group teaches nothing and falls silent. The same result says which problems deserve the most weight and how many tries each one needs. This paper confirms the intuition on a large real difficulty dataset (Big-Math) and in a controlled training run. What looks like a harmless normalization step is the dial that decides where learning happens and how strongly.

Online hate speech has been linked to a global rise in violence against minorities, including incidents such as mass shootings, lynchings, and ethnic cleansing. Societies grappling with this issue, particularly when hate speech targets specific groups based on religion, race, ethnicity, culture, nationality, or migration status, face the challenge of balancing freedom of expression with the need for effective content moderation on widely used online platforms. In response to this challenge, we introduce a comprehensive hate speech dataset covering five distinct topics in Turkish: refugees, the Israel-Palestine conflict, anti-Greek sentiment in Turkey, ethnic or religious communities (Alevis, Armenians, Arabs, Jews, and Kurds), and LGBTI+, alongside one topic in Arabic (refugees). In addition, we develop state-of-the-art BERT-based models to address multiple dimensions of hate speech analysis, including hate category classification, hate intensity prediction, target identification, and hate speech span detection, enabling a comprehensive understanding of hateful content in online discourse.

As computing education expands beyond traditional programming into operational domains such as systems administration and command-line environments, existing pedagogical frameworks struggle to capture a dimension that is critical in these contexts: the real-world consequences of learner actions. Existing cognitive taxonomies classify learning objectives by mental operations but do not account for system impact, leaving a critical gap in command-line education where conceptually simple commands can have severe consequences. This work presents CogTax, a four-level cognitive taxonomy that integrates two dimensions: cognitive complexity, derived from Bloom's Revised Taxonomy, and operational impact, which distinguishes observational, reversible, structural, and administrative operations. The four progressive levels range from safe read-only inspection to advanced system management requiring integration of multiple abstract models. Then, the taxonomy level is defined as the maximum of these dimensions, ensuring that both conceptual understanding and operational awareness are addressed. CogTax gives instructors a principled framework for sequencing course material and calibrating assessment difficulty, and gives students an explicit reference for self-assessment and gap identification. To demonstrate that taxonomy levels are automatically assignable, making the framework scalable without manual expert annotation, a classifier that combines syntactic representations derived from abstract syntax trees with semantic embeddings is trained. Evaluated on 585 expert-annotated Linux/bash commands, this combined approach achieves 89% accuracy, outperforming either representation alone, and demonstrates cross-language extensibility through structural equivalences across command languages.

The cost of human expert evaluation is a principal bottleneck to deploying language models in specialized, high-stakes domains. This is particularly acute for Arabic sociolinguistic knowledge: credible grading requires not only linguistic fluency but deep cultural familiarity that cannot be approximated by surface-level metrics. We address this with a cross-evaluation framework instantiated on two underrepresented Arabic dialect communities: Egyptian and Iraqi Arabic. We contribute 103 validated prompt-rubric pairs (70 Egyptian, 33 Iraqi; 53 Cultural, 50 Linguistic), authored and graded by native-speaker SMEs using penalty-weighted rubrics distinguishing positive content requirements from answer-specific negative error criteria. Three frontier LLMs serve as target models (graded by human SMEs across 302 unique prompt-response pairs), while five frontier LLMs serve as automated judges enforcing a provider-level self-evaluation guard. A dual-metric scheme combining Mean Absolute Deviation (MAD) with Signed Mean Error separates directional grading bias from symmetric noise. Across 1,307 judge evaluations: GPT-5.4 is the most reliable judge (MADj = 10.21 pp, Signed Error = -1.12%); four of five judges show systematic leniency (+2.01% to +6.56%); Cultural tasks are harder to grade than Linguistic tasks for all judges (MAD gap 1.83-4.78 pp); and models substantially outperform on Egyptian prompts compared to Iraqi prompts. However, given leniency differences between Iraqi and Egyptian SMEs, we cannot solely attribute this gap to model knowledge. We therefore emphasize findings that do not assume identical leniency across human graders. Across all samples, implicit cultural reasoning – requiring models to simulate native-speaker judgment rather than rely on lexical verification – emerges as the primary failure mode for automated grading across all judge models.

Language models have changed from unreliable text generators to highly-capable large models with trillions of parameters. Capability increases come hand-in-hand with increases in scale, making understanding the internal representations of models more challenging. Since millions of users increasing rely on language models to interact with external tools or make decisions in medium or high-stakes scenarios, we need to establish control over model behavior and know when to trust model outputs. In this paper, we discuss our contributions on harnessing the latent spaces by proposing steering vectors for control and developing latent space-based model calibrators for trust. Together, our contributions help demystify the latent spaces of language models and offer new insights into how to harness model internals to build more trustworthy language technology.

When does training language models (LMs) to generate explanations of their predictions yield faithful introspection, rather than superficial imitation? We study LMs trained to explain which features of their inputs influenced their behavior, using models' counterfactual behavior on modified inputs as supervision. Surprisingly, we find that LMs trained on fixed counterfactual explanations derived from earlier checkpoints of themselves, or even from behaviorally similar models in different families, frequently produce explanations more faithful to their own current behaviors than to those of their training targets. This "introspective" coupling between LM explanations and behaviors occurs when training explanations remain sufficiently correlated with current behaviors over the course of training, even as behaviors themselves shift. We also show that introspective coupling tracks behavior shifts: when explanation training is provided concurrently with other post-training objectives, explanations track those shifts without requiring updated supervision. This phenomenon appears in multiple tasks, including sycophancy and refusal, and is robust to label noise. Overall, our results show that even fixed datasets of counterfactual explanations can provide scalable and generalizable post-training signal for introspection.

LLM agents increasingly act over long horizons, where a single trajectory can contain hundreds or thousands of actions. In these settings, outcome-only rewards provide too sparse guidance, failing to inform the model about the goodness of intermediate actions. Dense supervision methods aim to solve this problem by scoring intermediate steps, from intrinsic confidence to self-distillation and embedding similarities. However, it is common practice to evaluate them by measuring the downstream performance of a training pipeline that integrates them. This is expensive, conflates supervision quality with training engineering confounders, and renders different methodological families requiring distinct training setups incomparable. As a result, dense supervision methods are rarely benchmarked on common ground. We introduce QVal, a training-free testbed for directly evaluating dense supervision signals. Given a state-action pair, QVal measures how well a method's score is Q-aligned: whether it orders actions according to the Q-values of a strong reference-policy. This lets us compare signals before any training run and separate signal quality from other engineering choices. We instantiate QVal as QVal-v1.0, benchmarking 21 dense supervision methods across four diverse environments and seven methodological families, with over 1.2K evaluation experiments across six open-weight model backbones. We find that simple prompting baselines consistently outperform recent dense supervision methods from the literature, and that performance clusters strongly by family. These findings hold across model sizes, environments, and observation modalities. QVal is designed to be easily extensible to new environments and methods, enabling researchers to iterate on dense supervision methods before any training run.

Metacognition is a critical component of intelligence that describes the ability to monitor and regulate one's own cognitive processes. Yet LLMs exhibit systemic deficiencies in key metacognitive faculties: they hallucinate with high confidence, fail to recognize knowledge boundaries, and misrepresent their internal uncertainty–undermining trustworthiness and reliability. Since monitoring task performance and adapting behavior accordingly are central to metacognition, we posit that models capable of accurately judging their own performance are better positioned to improve it. We operationalize this idea via two novel mechanisms: reinforcement learning with metacognitive feedback (RLMF), a paradigm to refine completion rankings during preference optimization based on the quality of a model's self-judgments of performance, and metacognitive data selection, which uses similar self-judgments to identify high-value training examples, outperforming naive active learning. We apply these innovations to the problem of faithful calibration (FC), a task that is itself fundamentally metacognitive: the goal is to align expressed with intrinsic uncertainty, difficult even for frontier LLMs. We adopt a two-stage, decoupled approach, first using these methods to calibrate the faithfulness of models' self-reported confidence scores, then mapping to natural, context-adaptable linguistic uncertainty via targeted output editing. Extensive experiments show RLMF achieves generalizable, state-of-the-art FC on diverse tasks while preserving accuracy. Further, RLMF surpasses standard RL by up to 63% while enhancing models' ability to assess and express their own capability limits. This positions RLMF as a promising paradigm to enhance LLM metacognition toward improved abilities and alignment, and suggests metacognitive performance as an effective RL signal to overcome limits of prior intrinsic feedback methods.

While large language models (LLMs) perform well on table tasks, they still make data referencing errors (DREs), i.e., incorrectly citing or omitting table values, despite understanding the table structure. Beyond final-answer accuracy, DREs directly compromise the correctness and reliability of intermediate reasoning steps. Yet prior studies have only offered limited, small-scale analyses. In this work, we present the first systematic evaluation of tabular data referencing errors across different models and tasks. Our results show that DREs occur across all tested models (1.7B to 20B parameters). Furthermore, we demonstrate that incorporating data referencing as a critic significantly improves answer accuracy up to 12.0%, through critic-based filtering and rejection sampling. Finally, we trained a lightweight 4B-parameter critic model that achieves an average F1 score of 78.2% in detecting both in-distribution and out-of-distribution DREs, and effectively assists inference for larger models.