Papers

Using large language models (LLMs) to assist psychological counseling is an important task in the field of natural language processing. The construction of high-quality psychological support dialogue corpora serves as a critical foundation for training counseling-oriented conversational models. However, existing data generation approaches generally suffer from several limitations, including emotionally stable seekers, limited variation in emotional dynamics, and a high degree of compliance with counselors' guidance. These issues result in LLM that lack the capability to effectively respond to emotionally unstable scenarios. In addition, counselor responses are typically driven by problem-solving objectives, thereby overlooking the role of emotion-focused interaction, which are essential in psychological counseling. To address these gaps, we propose EmoTrace, a multi-turn dialogue corpus generation framework centered on modeling seekers' emotional trajectories. we construct seekers' cognitive profile and introduce a seeker module with emotional schemas and an associated activation mechanism, a counselor module, and an emotional trajectory control module, thereby enhancing the layering of the seeker's emotional expression and the counselor's targeted empathic expression. Experimental results demonstrate that the proposed method outperforms existing approaches in terms of emotional richness and empathy quality.

Third-party API routers have become a common layer that unifies access across increasingly diverse LLM providers. In coding-agent workflows, high-autonomy operation is widely adopted because it reduces interaction overhead. As a result, a third-party API router, which sits between the agent and the upstream provider, inevitably occupies the trusted path. It can inspect and modify every request and response, yet no mechanism verifies alignment between the provider's output and the repository-level actions ultimately executed by the agent. Consequently, client-side permission mechanisms may become ineffective in practice. Whether this control gap produces real, hard-to-detect effects on software development tasks remains empirically unmeasured. In this paper, we conduct an empirical study of router-side injection in coding agents, examining four intervention levels of increasing subtlety: Response Substitution (L1), Response Append (L2), LLM-Polished Injection (L3), and LLM-Polished with Distribution Alignment Injection (L4). Moreover, we develop SIDEL, a framework for trace recording, replay, injection, and defense evaluation, with a curated dataset of 400 samples. We evaluate four representative coding agents, and further evaluate whitelist-based execution control and LLM review. Router-side intervention substantially alters repository-level actions and remains difficult for existing client-side safeguards to detect. Without additional mitigations, all evaluated agents achieved a defense success rate of 0 percent across all injection levels. Client-side mitigations and reactive reviews improve resistance but do not fully restore end-to-end control, motivating provider-side output-integrity guarantees. Our code is available at https://github.com/Riyasushin/SIDEL.

This paper presents GEMCo, a releasable, human-written proxy for inaccessible counselling data: 86 complete German e-mail counselling conversations (728 messages), expert-authored cases and counsellor sessions with trained role-players. It is validated against a held-out reference of 124 real counselling conversations. The proxy and the real conversations are measured against each other on counsellor strategies and client emotions. The gap is detectable but small. A generative validation supports the analysis. The corpus is the primary contribution. The validation method generalises to any domain where real data cannot be shared but a human-made proxy can. Privacy and ethics keep real counselling data closed. GEMCo carries none by design and can be released, a first step toward language research in this domain.

Molecular structure elucidation from tandem mass spectra (MS/MS) is a central inverse problem in analytical chemistry. Most existing approaches to MS/MS identification remain tied to reference libraries or predefined candidate sets, whereas de novo methods aim to generate structures directly from spectra. A common de novo route predicts a molecular fingerprint from the spectrum and then decodes structures from it, enabling decoder pretraining on large molecule-only corpora. However, this paradigm creates a training-inference mismatch: the decoder is trained on oracle fingerprints computed from molecules, but at inference it is queried with a noisy spectrum-induced fingerprint posterior that is typically collapsed to a single thresholded fingerprint. We introduce MS-GPT, which recasts fingerprint-mediated de novo elucidation as spectrum-induced posterior querying of a conditional molecule-language model. MS-GPT conditions a molecule-language model on fingerprints and formulas, then converts the spectrum-induced posterior into a band of fingerprint queries near the oracle-fingerprint manifold through active-bit density calibration. Candidates sampled across this band are pooled and ranked by generation-frequency consensus. A lightweight LoRA adapter further mitigates domain-specific posterior bias while preserving the pretrained molecular prior. On NPLIB1 and MassSpecGym, MS-GPT sets a new state of the art, reaching Top-1/Top-10 exact-match accuracy of 29.8%/41.1% and 23.9%/28.7%, respectively. Candidate-pool scaling shows that efficient autoregressive molecular generation continues to improve recall with a little additional inference cost. The source code and model checkpoints are available at https://github.com/VIKI623/MS-GPT.

Generative linguistic steganography conceals secret bits within the sampling randomness of large language models. Existing schemes are single-stream, conveying an entire secret through a single response to a single prompt. This convention incurs limitations: it provides no protocol-level support for batched multi-stream inference, and naive co-batching does not conceal slot occupancy or payload completion. We propose the High-Throughput Multi-Stream (HiTMS) framework, which distributes a secret across multiple responses produced jointly over successive rounds of interaction. Each round embeds and extracts several streams within a single batched call, thereby amortizing the cost of model invocation and substantially improving throughput. To ensure recoverability, HiTMS wraps each response in a self-describing frame and employs a key-derived schedule that binds streams to slots and fills unused slots with decoys, guaranteeing exact recovery while concealing the number of active streams. The framework is agnostic to both the language model and the steganographic coder. Across eight dataset-model-coder settings, eight-stream HiTMS achieves up to 4.3 times higher embedding and extraction speeds than single-stream baselines, while reducing the average area under the receiver operating characteristic curve (AUROC) of steganalyzers from 0.681 to 0.601. Experiments with 4 to 64 streams demonstrate sustained throughput gains as concurrency increases. GitHub repository for this work is https://github.com/ryehr/HiTMS_steganography.

Instruction hierarchy (IH) requires models to prioritize instructions by source, ensuring that higher-priority instructions override lower-priority ones. Despite its importance for safe and controllable deployment, existing evaluations have focused almost exclusively on English, leaving it unclear whether IH compliance remains stable in multilingual settings. We introduce XIH-Bench, a benchmark for multilingual IH evaluation with both same-language and cross-language conflicts across six languages, four domains, and three IH settings. Across models, we find two consistent patterns. First, IH compliance exhibits a clear language-dependent asymmetry: a language that strengthens compliance in the higher-priority position can become disruptive in the lower-priority position. Second, cross-language conflicts yield higher compliance than same-language conflicts, a phenomenon we term the Language Boundary Effect. We further show that language specialization can make lower-priority instructions in model-favored languages harder to override, creating multilingual reliability and security risks.

Despite recent advances in large language models (LLMs), their ability to generate empathetic mental health counseling responses in low-resource languages remains largely unexplored. To address this gap, we curate 625 authentic mental health cases from three complementary sources: (1) publicly available Facebook posts discussing mental health concerns, (2) transcripts from the Bangladeshi television program "Ami Akhon Ki Korbo", and (3) anonymized student questionnaire responses covering diverse emotional and psychological challenges. Based on these cases, we build an evaluation corpus comprising advice written by licensed clinical psychologists and responses generated by three modern proprietary LLMs: GPT-4o Mini, Claude 4.5 Haiku, and Gemini 2.5 Pro. We further propose the Role-Playing Reflective Chain-of-Thought Advisory Framework (RP-RCAF), a task-specific prompting strategy that combines expert-authored few-shot examples with structured self-reflection to produce supportive, culturally aware, and ethically aligned counseling through a compassionate advisor persona. We also introduce the Grok 4-Based Response Evaluation and Scoring Framework (G-REFS), which integrates automated assessment with expert psychologist validation across emotional sensitivity, cultural appropriateness, linguistic clarity, and ethical soundness. Experimental results show that RP-RCAF consistently outperforms conventional prompting across all evaluated models and produces responses that more closely align with professional psychological counseling.

Joint-Embedding Predictive Architectures (JEPAs) are effective for images, video, and audio, yet deterministic JEPA-style latent prediction has not become a standard objective for text encoders. We argue that this gap reflects a mismatch between squared-error latent prediction and the conditional structure of language. The key requirement is conditional concentration: given a context and target location, the target representation should lie near a single meaningful point. Local image prediction often satisfies this through spatial continuity, whereas masked text can admit multiple valid token or span completions whose representations need not share a coherent center. We formalize this mismatch through three conditions—predictability, non-collapse, and low conditional variance—and show how their failure creates centroid degeneracy and collapse pressure in text. Matched I-JEPA and T-JEPA experiments reveal the predicted sequence: mutual-information saturation and elevated target variance precede train–validation instability, effective-rank degeneration, cosine collapse, and poor downstream transfer. The same pattern appears across five independent data seeds, indicating that it is not a sampling artifact. These results do not rule out predictive learning for language; they show that text-compatible JEPA objectives must preserve multiple plausible completions rather than compress them into a single latent point.

Political audits of large language models (LLMs) usually reduce each to one point on a political compass. But that resting point barely matters in deployment: a model must land somewhere, and what counts is how far, and in which directions, its answers can be steered. That steering runs through the system prompt: the personalization layer a platform sets, or one induced from a user's history, not necessarily written by hand. We run a dispersion-first stress test of prompt-based controllability across 12 ideological personas plus an unsteered baseline, 70 Political Compass items, ten replicates, and seven leading LLMs: GPT-5, Claude, Grok, Gemini, DeepSeek, Kimi, and Qwen (63,700 responses). Contextual framing explains roughly 88%-93% of variance on the economic and society axes, model identity under 3%: responses are highly instruction-adjustable. Models do not shift alike: some move more, and some saturate under extreme framings. Conflicting directional-steering results in prior audits resolve once baselines are recognized as non-centered: displacement and proximity diverge, so the effect is geometric, not differential compliance. Under authoritarian prompts, models produce similar shifts on the same questions. Political-coordinate audits therefore need steerability audits reporting dispersion, symmetry, saturation, and refusal floors. We release prompts, benchmark data, and code.

Multimodal automated fact-checking (MAFC) verifies claims by retrieving and reasoning over external evidence. However, most existing static benchmarks risk contamination: they primarily consist of outdated claims verifiable using an LLM's internal knowledge without external evidence. This can inflate performance estimates and fail to reflect true capability on novel claims that require up-to-date information. To address this, emerging dynamic benchmarks collect claims published after LLMs' knowledge cut-off dates, assuming they are uncontaminated. This work revisits this assumption by empirically studying contamination risks in both the state-of-the-art (SOTA) static AVeriTeC benchmark and our newly constructed dynamic ClaimReview2025Q4 benchmark, as well as their impact on MAFC evaluation. Our experiments yield 16 findings, highlighting three key results: (1) Dynamic evaluation reduces but does not eliminate contamination risks, as 17.09%–29.30% of post-cut-off claims remain potentially contaminated; (2) Many newly published claims can be verified either directly or by synthesizing multiple pieces of public knowledge available before the cut-off; and (3) Contamination can induce statistically significant inflation in MAFC performance, increasing Macro-F1 by up to 11.34 points and distorting system rankings. In light of these findings, we re-evaluate SOTA LLMs under a strictly contamination-controlled setting. Our study provides practical guidelines for trustworthy MAFC evaluation.

Diagrams are widely used to support logical reasoning, and prior studies suggest that representations such as Euler diagrams can improve human reasoning performance. Recent work has also explored their effects on large language models (LLMs). In this paper, we compare four representational conditions for syllogistic reasoning: natural language, logical notation, linear diagrams, and Euler diagrams. Using 285 problems from Ando et al. (2024), we evaluate two contemporary LLMs, Claude 3.5~Sonnet and GPT-4o-mini. Our results show that diagrammatic representations do not consistently improve performance. Although the models perform well on entailment and contradiction problems, they struggle with neutral problems and often make systematic conversion errors. Overall, the results suggest that the tested models gain limited benefit from diagrams in logical reasoning tasks.

Offensive language detection models generally suffer performance degradation when deployed across datasets and across languages, yet most existing studies stop at reporting this phenomenon and lack a systematic methodology for decomposing the causes of degradation into attributable components and quantifying the cost of remediation. This paper proposes a diagnosis and optimization framework composed of three coordinated technical components. First, a zero-shot transfer loss decomposition that separates the performance degradation from OLID to MLMA into two independently measurable components, namely dataset effect and language effect. Second, a controlled fine-tuning protocol that quantifies both adaptation efficiency and the hidden damage inflicted on the source task by comparing few shot learning curves under continued fine-tuning and cold-start starting points. Third, three joint training strategies incorpo rating temperature sampling and experience replay, which offer a controllable Pareto trade-off between improving multilingual capability and preserving source-task performance. Experiments built on this framework show that the dataset effect dominates the zero-shot transfer loss and substantially outweighs the language effect. Few-shot adaptation without a replay mechanism, though data-efficient, inflicts source task damage 4 to 9 times greater than that of the joint training strategies, and its damage magnitude is highly unstable. The three joint training strategies trade 3.2 to 4.1 percentage points of source-task performance for 8.1 to 42.6 percentage points of multilingual capability gain, forming a clear and controllable Pareto trade-off.

Two structural insights have been overlooked in automated residential floor plan generation. First, design is inherently progressive. Architects begin with rough strokes and refine them over time, whereas existing methods typically require their conditioning representation to be fully specified before generation, a fundamental mismatch with how design actually works. Second, the 2D floor plan is not an optional intermediate but an irreplaceable spatial contract. Once room boundaries, doors, and windows are fixed, furnishing reduces from open-ended spatial reasoning to bounded constraint satisfaction. Bypassing this contract, as existing 3D systems do by delegating layout to language models, yields overlapping rooms and implausible proportions; directly calling general-purpose language models likewise produces geometrically invalid layouts. Guided by these insights, we present PlanCraft. SketchPlan supplies the missing training signal by replaying the architect's drawing process on 80K real floor plans, producing partial sketches at every completeness level. PlanCraft-Diff progressively sharpens an incomplete sketch into a geometrically precise, vectorizable floor plan through a coarse-to-fine strategy. With the spatial contract established, PlanCraft-Agent then furnishes the scene within well-defined room boundaries. Experiments show that PlanCraft achieves a 61.1% lower FID than the best existing 2D method and surpasses existing 3D systems by 15 points in expert-rated spatial rationality, with a sketch at only 25% completion already outperforming all fully specified baselines.

This paper presents Mwando, a virtual educational assistant designed to support the teaching and preservation of shiKomori, the language of the Comoros Islands. The system covers the four main dialectal variants (shiNgazidja, shiMwali, shiNdzuani and shiMaore) through a knowledge base constructed from phrases, proverbs, dictionaries and grammar lessons. A multi-agent architecture combining vector search, a knowledge graph and web search fallback enables accurate and context-aware responses. Evaluation on 500 queries demonstrates strong performance on vocabulary lookup and grammar explanations, while qualitative case studies illustrate both capabilities and current limitations. This work represents an initial step toward computational support for shiKomori and provides a blueprint for developing AI-powered educational tools for other low-resource languages.

Chain-of-thought (CoT) explanations support oversight only if they are faithful: the stated reasoning must actually produce the answer. Auditing black-box (behavioral) detection of unfaithful CoT against FaithCoT-Bench's human annotations, we find answer correctness structures the problem at every level. Answer incorrectness alone (an oracle diagnostic, not a deployable detector) outperforms every purpose-built signal (AUROC 0.696), because 69% of annotated unfaithfulness occurs on incorrect answers. Stratifying by correctness splits detection into two regimes: on correct answers, behavioral signals moderately separate faithful from post-hoc reasoning (0.63-0.67); on incorrect answers, where most unfaithfulness lives, no tested signal is detectably above chance (replicated on all four models for benchmark-wide signals). The standard step-removal metric anti-correlates with human labels; this inversion reproduces on the benchmark's released scores and on hint-dependent counterfactually labeled traces. Linear probes decode the behaviorally blind regime in Llama-3.1-8B and the correct-answer regime in Qwen-2.5-7B, with no shared, positively aligned direction detected across regimes; instructed answer-first traces (7 models) transfer to neither annotated regime, while hint-induced unverbalized answer flips do, in model- and source-dependent settings. We also independently verify and resolve a documentation-data mismatch in the benchmark's label semantics.

A small language model can receive the governing statutory provision and still answer incorrectly. We test whether fine-tuning on examples containing relevant law improves later use of retrieved law. We curate 2{,}165 bilingual QA records from six Bangladeshi acts and three schedules, then fine-tune Qwen3.5 at 0.8B, 2B, and 4B. Evaluation uses the 2022 and 2023 Bangladesh Bar Council exams in Bangla and machine-translated English, with no retrieval, BM25, or FAISS, scored by strict consistency over three seeded runs. At 0.8B, fine-tuning raises the 2022 English FAISS score from 2 to 34 of 100. Gains at 0.8B and 2B survive paired testing, but the 4B model has no detectable net gain: Bangla improves while several English conditions regress. Fine-tuning also reduces answers that drift from Bangla into mostly English from 44.0–53.2% to 0.2–0.7%, with adjusted $p<.001$ at every scale. Retrieval quality is therefore not the only bottleneck. Small bilingual legal models also differ in how they use supplied law and whether they answer in the requested language. The dataset is publicly available at https://huggingface.co/datasets/momahadi/bangladesh-legal-qa-dataset.

Omnimodal large language models (OmniLLMs) are rapidly extending multimodal reasoning to cover synchronized audio and video. However, the resulting audio-video token sequences are long, leading to high prefill latency and GPU memory usage at inference time. Existing token pruning methods, designed mainly for vision-only inputs, miss both the cross-modal links between audio and video and the user query that decides which content matters. To bridge this gap, we present Omni-Prune, a training-free, query-aware audio-visual token pruning framework that jointly removes redundancy from both modalities while keeping task-relevant cross-modal evidence. Specifically, Omni-Prune first splits the token sequence into adaptive time windows placed at audio saliency peaks, then scores audio and video tokens on a single scale that combines encoder attention with text-query relevance, and pairs related audio-video tokens so that they are kept together. Within each window, a final K-medoids step then selects a few representative tokens, adding diverse cues that score-based selection alone would miss. Extensive experiments demonstrate that Omni-Prune outperforms established baseline methods, delivering up to 3.25x prefill speedup and 1.3x memory reduction while retaining over 99% of full-model performance.

LLM debate is usually evaluated by final answers, yet transcripts reveal whether later turns develop new arguments or return to earlier claims in new wording. We study this process with prior-argument similarity, which compares extracted argument units with earlier units in the same debate. In controlled eight-turn debates over 71 motions, six languages, and four model agents, Chinese is the only tested language with a consistently positive gap relative to English across three multilingual embedding models. The gap persists across agents, turn positions, regression adjustment, metric variants, extraction-length controls, a second-extractor subset, and cross-encoder tail rescoring. Manual calibration shows weak item-level alignment but a high-similarity tail enriched for substantive repetition. A diversity-aware prompt lowers prior-argument similarity across languages, yet does not significantly narrow the Chinese–English gap. Multilingual debate evaluation should therefore measure argumentative development over time and report both average and gap terms.

In this paper, we push the boundary of LLM reasoning by testing them in a Chinese language game, xiehouyu, with novel xiehouyu created by linguists that had not existed before to avoid data contamination. We use multiple-choice questions (MCQ), free-form explanation generation, and new xiehouyu creation to evaluate LLMs' ability to understand and create xiehouyu. In MCQ, we use the delta of accuracy ($\Delta_{acc}$) between existing but low-frequency xiehouyu and novel ones as an index for memorization. $\Delta_{acc}$ for native speakers is very low, suggesting similar processing mechanisms. However, we found that frontier Chinese models have on average a $\Delta_{acc}$ of 23.6%, while English-centric models tested have a mean $\Delta_{acc}$ of 5.1%, suggesting that frontier Chinese models are likely trained with much larger Chinese data, thus memorizing more low-frequency xiehouyu. For novel xiehouyu, Gemini 3.1 Pro demonstrated remarkable ability with acc 92.6, which is 24% higher than human accuracy. In xiehouyu creation, those created by LLMs receive much worse ratings than those by humans. These results suggest that claims about the reasoning abilities of LLMs may need careful re-examination considering the data contamination issue, and that LLMs' creativity in language-related tasks may still be behind human experts, at least in Chinese xiehouyu.

Large language models show strong promise for information extraction (IE), but existing reflection-based correction methods are often misaligned with structured extraction outputs. Free-form self-reflection can flag an error, yet it rarely identifies whether the failure is a missing span, wrong label, boundary mismatch, invalid relation type, or reversed argument order. We introduce LA-RL (Label-Aware Reflective Reinforcement Learning), an outcome-supervised framework that guides IE self-correction with task-grounded diagnostic labels. A single backbone first predicts an extraction, diagnoses task-specific error labels, and then revises its output conditioned on the diagnosis. Training starts from diagnostic data labeled by an annotation model for cold-start supervised fine-tuning and proceeds through two GRPO stages that reward final extraction quality, format validity, and first-pass correctness, without a process reward model. Experiments on named entity recognition, relation extraction, and event extraction show consistent same-backbone gains over SFT, including 6.83 average F1 on SciER relation extraction, about 20 F1 on out-of-distribution relation extraction, and 14.80 trigger F1 plus 17.50 argument F1 on DuEE1.0. Ablations show that reflection structure is task-sensitive: stronger constraints benefit relation extraction, whereas named entity recognition needs less restrictive correction under domain shift.

In the rapidly evolving landscape of cloud-native computing, Organizations are increasingly adopting infrastructure models that emphasize scalability, flexibility, and efficiency. Kubernetes has become the de facto standard for orchestrating containerized applications in these environments. However, the inherent complexity of cloud-native ecosystems introduces significant challenges, particularly in the form of misconfigurations that can compromise both security and performance. This study explores the potential of Large Language Models (LLMs) in identifying Kubernetes misconfigurations. We introduce a comprehensive taxonomy of common misconfiguration types, offering a structured framework to better understand and categorize these issues. Additionally, we conduct an empirical evaluation of state-of-the-art detection tools to benchmark their effectiveness. Furthermore, we analyze the Kubernetes objects most prone to misconfiguration and evaluate the severity of the identified issues. By leveraging advanced machine learning techniques, including LLMs, we provide novel insights into enhancing misconfiguration detection methodologies.

Large language models produce prompt-dependent probabilities over words, whereas scientific systems require uncertainty over meaningful states that can be updated as evidence arrives. We develop an observable framework for determining when language-derived probabilities support such a sequential state representation. Theoretically, we define typed measurable transformations of contextual language, construct a minimal closed representation, and give necessary and sufficient conditions for semantic updates to exist uniquely. We bound irreducible nonclosure and accumulated error, and under average contraction prove existence, uniqueness and stability of an external random recursion on a probability simplex. These results define a stochastic lexical calculus without attributing an internal calculus to the language model. Empirically, frozen experiments test the observable implications. Raw prompt-conditioned probabilities fail the prespecified invariance gate; after prompt-specific calibration, a common three-state representation passes the stability gates and covers 28 of 30 untouched eight-step paths, or 0.933 at nominal level 0.90. Accordingly, language probabilities support a stochastic state only conditionally on verified closure, stability and coverage within a declared operating domain.

Materials research advances through accumulated experience - scripts that work, protocols that are trusted, warnings attached to failed calculations or experiments, and judgement that links a new question to an old result. This experience is essential for reproducibility and knowledge transfer, yet it is usually fragmented across notebooks, repositories, job logs and individual memory, and it is rarely portable across artificial-intelligence agents. Here we argue that a lifelong AI partner for materials science can be designed around persistent memory rather than around a particular agent implementation. We introduce a self-evolving memory framework that stores scientific experience as inspectable facts and executable skills, so that observations, failure boundaries, protocols and validation checks can be retrieved, revised and migrated across models. We evaluate the idea in three computational settings that expose different layers of materials-research competence. In 49 real-world materials-tool-use questions comprising 138 executable subtasks, memory nearly doubles GPT-5.2 task success without model-parameter updates. In elemental-solid equation-of-state calculations, memory converts a wavefunction-initialization failure into a pre-execution guardrail, improving outcomes from 22/1/4 to 25/2/0 Correct/Partial/Error and avoiding 92% of repeated errors. In 13 practical material simulation workflows, remembered skills and failure facts halve the aggregate trace burden (tokens) and reduce tool calls by over a factor of two by the third round, while preserving physically meaningful outputs in band-gap, phonon, vacancy and work-function analyses. These results show that agent memory can serve as a durable scientific asset; a portable, self-improving record of materials-research experience that outlives any single model or agent stack.

Enterprise AI programmes stall at a rate that is widely quoted and poorly explained. This paper measures the mechanism. Six document-heavy workflows of the kind performed daily in regulated financial services were run across four model families and three tool configurations, three times each, producing 5,093 scored output elements across 72 configurations. Each configuration was assessed twice: against a demonstration bar, being a single correct run on a single case, and against a production bar requiring sustained accuracy, reproducibility across repeats, verifiable attribution, and a confidence signal that carries information. 57 of 72 configurations cleared the demonstration bar and 32 cleared the production bar, a survival rate of 56.1%. The paper then computes the review burden each configuration imposes, estimated out of sample rather than with hindsight. A tool that states no confidence requires review of 100% of its output, because it offers a reviewer no basis for triage. Requiring the tool to cite its sources and state a confidence reduces that to 49% while holding the residual error tolerance in 17 of 20 configurations. Adding a self-verification pass costs 2.3 times the latency of the plain configuration, reaches 44%, and is the only configuration that fails to hold the error tolerance. The practical implication is that the value of an AI workflow is set less by how often it is right than by how much of it a human must still check, and that the second property is measurable and rarely measured.

Activation Oracles (AOs) are language models trained to answer natural-language questions about another model's internal activations. They offer a flexible interface for reading hidden information from model states, especially when relevant information is internally represented but absent or incomplete in visible behavior. However, AOs are themselves learned systems: their answers are shaped by training data, objectives, and learned reporting behavior, rather than being neutral readouts of represented information. We study this in a controlled Taboo Word Guessing setting, where subject models are fine-tuned to internally use a hidden concept while avoiding direct disclosure. Contrary to the expectation that an AO trained on such a subject becomes a specialist reader, we find that fine-tuned AOs can become concept-specific anti-readers: they selectively fail to recover the concept persistently present during their own training. This failure is not simply explained by absence of the concept from the subject or oracle representations: the target remains decodable inside the oracle, while LogitLens and layer-ablation analyses indicate that the failure arises in the AO readout pathway. Our results show that behavioral leakage, representation-level decodability, and AO-verbalizability can come apart, raising a reliability concern for learned interpretability interfaces.

Vision-Language Models (VLMs) are demonstrating significant capabilities in medical tasks like radiology analysis, yet providing faithful and interpretable explanations remains a key consideration for their responsible deployment in clinical settings. However, existing explanation methods, such as the widely used FIxLIP framework, often struggle with the fine-grained nature of modern tokenizers. The tokenization problem fragments clinical concepts—splitting terms like "saddle embolus" into scattered, meaningless subwords—which leads to noisy, semantically incoherent cross-modal attributions. Such fragmentation also results in a combinatorial explosion of interaction possibilities, obscuring the model's true reasoning. To address this, we introduce ParseFIxLIP, an extension that incorporates the Tree-Gram Parsing into the Banzhaf interaction game used by FIxLIP. This semantically informed strategy utilizes dependency parsing trees to define explanation players by grouping related text tokens into semantically coherent units. Our smart_depth grouping strategy, merging tokens according to spaCy token dependency tree, successfully mitigates concept fragmentation, yielding substantially more interpretable cross-modal interactions by unifying complex medical concepts. Quantitatively, while baselines struggled with the high dimensionality of long captions, our parsing approach maintained statistical robustness and semantic parsimony. Qualitative analysis on BiomedCLIP, validated on medical imagery (ROCOv2) and general examples, confirms that the approach accurately captures the synergistic influence of grouped words on model predictions. In conclusion, our work offers intuitive and clinically relevant insights into VLM decision-making, fulfilling the critical need for coherent explanations in the medical domain.

Recent work has shown that subword vocabularies can be trained to optimize compression for a specific inference rule rather than relying on greedy heuristics such as Byte Pair Encoding (BPE). We extend this approach to greedy left-to-right longest-match decoding, the fast and widely used inference rule underlying WordPiece. We introduce Joint Optimization for Greedy Longest-Match Tokenization (JOLT), which formulates vocabulary learning as an integer program over vocabulary-selection and segmentation-choice variables. Greedy-consistency constraints ensure that each optimized segmentation exactly matches the segmentation produced by longest-match decoding under the selected vocabulary, aligning the training objective with deployment-time tokenization. To scale the optimization, we solve a linear programming relaxation and selectively introduce higher-order segmentations only for unresolved pretokens. The resulting relaxation is nearly integral: rounded solutions fall within 0.008 - 0.176 % of the LP lower bound on the training scope. The bound also shows that BPE is already within 1 - 2 % of the best achievable compression under greedy longest-match decoding, while JOLT closes 89.6 - 99.4 % of the remaining gap. On held-out validation data across four training scopes and vocabulary sizes of 32,000 and 64,000, JOLT produces up to 0.78 % fewer tokens than BPE, with improvements generally increasing as the training scope grows. These results demonstrate that inference-aligned vocabulary optimization can recover most of the limited compression headroom left by BPE while providing a certificate of near-optimality.

Language generation in the limit is an elegant model introduced by Kleinberg and Mullainathan [KM24] to formally study language generation by an algorithm that learns solely based on example strings. In this model, an algorithm is said to correctly generate from a language if it never makes an error after some finite time. In contrast, even sophisticated language models are known to regularly hallucinate in practice. In this paper, we initiate the study of language generation in the limit with (infinite) hallucination, i.e., the algorithm may generate incorrect strings infinitely often, but the errors occur at a limited rate (possibly even with 0-measure). We first show that hallucination, even at rate 0, makes generation in the limit strictly more powerful: there are language collections that cannot be generated with finite error but can be generated with infinite error, even when errors occur on a 0-measure set of time-steps. Furthermore, while all countable collections are generatable with finite error, we show a strict hierarchy of (uncountable) language collections characterized by the hallucination rate. This hierarchy extends to breadth, the fraction of the target language generated. While all countable collections can attain the optimal breadth of 1/2 [KW26b], we show strict separation at every breadth and hallucination rate. Finally, we study generation in the limit without repetition, where the algorithm may not repeat strings. This lets us compare the sets of correct and incorrect strings generated, rather than the fractions of correct and incorrect time-steps. Once again, we demonstrate a strict hierarchy at every hallucination rate and breadth. Taken together, these results reveal rich structure in language collections generatable in the limit with hallucination and establish hallucination rate as an important parameter in the theoretical study of language generation.

Named Entity Recognition (NER) for low-resource languages such as Marathi remains a challenging task due to limited annotated resources and linguistic complexity. Although recent Large Language Models (LLMs) have demonstrated strong performance across a wide range of natural language processing tasks, their effectiveness for language-specific NER in low-resource settings remains uncertain. In this study, we fine-tune MahaBERT-v2 on different variants of the MahaNER dataset and systematically compare the performance of these models with an existing MahaNER baseline and prominent general-purpose LLMs, including Gemini, LLaMA-3.3-70B, and Gemma models. All models are evaluated on a Marathi NER test dataset using standard metrics of precision, recall, and F1-score. The experimental results show that the fine-tuned MahaBERT-based models consistently outperform both the baseline and all evaluated LLMs, with the fine-tuned models achieving F1-scores ranging from 0.88 to 0.91, surpassing the existing MahaNER model (0.8843) and significantly exceeding the performance of LLM-based approaches, whose F1-scores range from 0.57 to 0.69. These findings demonstrate that task-specific, language-focused models trained on domain-relevant data remain more effective than general-purpose LLMs for Marathi NER, highlighting the continued importance of specialized architectures for low-resource language processing.

Instruction tuning has become the standard method for adapting large language models to follow human intent, yet existing instruction datasets are dominated by English-language general-knowledge tasks and lack coverage of specialized pedagogical domains. This paper presents IKS-Instruct, a dataset of 24,795 instruction-response pairs for teaching language models to deliver educational content grounded in Indian Knowledge Systems (IKS). The dataset spans seven languages (English, Hindi, Sanskrit, Tamil, Telugu, Kannada, and Malayalam), covers 41 pedagogical techniques from the Vedic oral and mathematical traditions, and is aligned with the Central Board of Secondary Education (CBSE) curriculum for classes 6 through 12. The pairs are derived from six source types: classical text corpora (Bhagavad Gita, Thirukkural, Sangam literature, Vedic texts), curriculum-aligned pedagogical templates, Vedic mathematical sutra demonstrations, bilingual instruction pairs, technique-grounded multi-turn dialogues, and cross-tradition comparative analyses. Quality is assessed through a multi-judge evaluation framework in which independent language models score responses on 12 dimensions including technique fidelity, pedagogical quality, factual accuracy, and IKS cultural depth. Under a uniform five-judge external panel (median aggregation over 1,201 stratified items), the strongest IKS-Instruct fine-tune of a compact 7B model reaches a median judge score of 6.39, within 0.15 of a strong general-purpose reference model (Nemotron-Nano at 6.54) at a fraction of its deployment cost, while the base model without IKS fine-tuning scores near zero on the IKS-specific dimensions. Model quality does not increase monotonically with data curation, a result we report together with the corresponding data-quality gains.