Papers

Memory for a long-horizon LLM agent is a contract about what each future decision is allowed to see. The simplest contract appends past observations, tool calls, and reflections to every prompt, which makes prior context easy to access but also turns it into a jumbled mixture in which the effect of any single memory component is hard to isolate. We introduce and instrument an alternative bounded contract: every decision is made from a fresh user message assembled by typed retrieval, with no raw cross-decision transcript appended. The prompt thus stays bounded across runs of any length, and any single layer can be ablated in isolation. We instantiate the contract in Slay the Spire 2, a closed-rule stochastic deck-building game whose runs require hundreds of tactical and strategic decisions. A public online benchmark of frontier LLMs on the same game reports zero wins at the lowest difficulty across five configurations, and the developer-reported human win rate at the same difficulty is 16%; the task is hard but not saturated. Within our harness, a fixed-A0 ablation shows the largest observed difference when triggered strategic skills are enabled: the no-store baseline wins 3/10 games and adding the skill layer 6/10. At this sample size the comparison is directional rather than statistically decisive (Fisher exact p\approx0.37); a cross-backbone probe and public accumulating-context baselines are reported as operational comparisons rather than controlled tests of the contract variable itself. We release a reproducible testbed: 298 completed trajectories with condition tags, frozen memory/skill snapshots, prompt records, and analysis scripts – an agent design and a validated, reusable methodology for studying how explicit memory layers shape long-horizon LLM-agent decisions.

LLM-as-a-Judge has become the dominant evaluation paradigm for many natural language generation tasks (albeit mostly in English) due to shortcomings of conventional metrics and high correlations with human judgment. There are now attempts to extend LLM-as-a-Judge to multilingual settings including low-resource languages. However, LLMs have limited proficiency in low-resource languages, and there is often no adequate human validation in these settings. To highlight the scope of the problem and current practices, we explore the use of LLM-as-a-Judge evaluators in ACL Anthology papers focusing on multilingual settings and low-resource languages across a diverse set of tasks. Out of 650 papers mentioning LLM-as-a-judge, only 33 of them focus on low-resource or multilingual settings. Our in-depth analysis of these papers indicates inconsistent evaluation outcomes, a tendency to overtrust LLM judgments in multilingual settings, and the widespread reliance on a single judge model per study. To help the NLP community further, we conclude with a checklist of recommendations about how to use LLM-as-a-Judge in multilingual and low-resource settings

Instruction tuning for speech language models (SLMs) is substantially more challenging than for text-based large language models (LLMs), as it requires learning a new modality and a wide range of speech-specific instructions in addition to those supported by text LLMs. Existing SLM training approaches largely replicate the text LLM training paradigm by synthesizing large-scale speech pre-training and instruction-tuning datasets. However, this strategy is difficult to scale, since speech sequences are significantly longer than text sequences. In this paper, we propose SpeechCombine, an instruction-following speech language model trained without any instruction tuning, using only a single round of speech pre-training on 30k hours of data. Starting from a text LLM base model, we perform continuous pre-training on speech utterances to obtain a speech-adapted model, and then directly combine its weights with the weight difference between the instruction-tuned and base versions of the text LLM. Our results show that this simple combination strategy not only preserves the knowledge and capabilities of the original text LLM, but also effectively transfers them to the speech domain. These findings suggest a new direction for SLM training that avoids reliance on massive speech data.

Large language models (LLMs) exhibit remarkable reasoning capabilities, but their task-specific fine-tuning is notoriously plagued by overconfidence, severely hindering trustworthy deployment. We propose Data-Adaptive Lower-Rank Adaptation (DALorRA), a simple and effective variational Bayesian sparse framework that shifts the paradigm of uncertainty quantification from the dense parameter space to the lightweight rank level of low-rank adaptation (LoRA). With the insight that LoRA essentially aggregates multiple rank-one components that may provide superfluous model capacity, DALorRA imposes stochastic masking on rank dimensions, enabling Bayesian regularization of model capacity during training and ensemble-like calibration during inference. Extensive experiments demonstrate DALorRA's excellent calibration of LLMs without compromising reasoning accuracy.

We present HaloGuard 1.0, an open-weights implementation of the constitutional-classifier paradigm for input safety. It achieves state-of-the-art performance on English and multilingual prompt-safety benchmarks at roughly one-tenth the model size of current leading open guard models. The safety constitution is the organising structure of the corpus: a natural-language constitution of 46 policies and 2,940 subcategories drives synthetic data generation, with exhaustive one-to-one paired counterfactuals that hold topic and vocabulary fixed while flipping intent, a two-tier harmless design that separately targets boundary and baseline false positives (FPs), and balanced multilingual materialisation across 46 languages that treats language as a surface form appearing on both sides of the boundary rather than as an adversarial signal. Across seven prompt-safety benchmarks, HaloGuard 1.0-0.8B attains the best average F1 (90.9) of any open guard we evaluate, outperforming baselines up to 27B parameters (over 30 times larger) while holding false-positive rate (FPR) to 4.3 and false-negative rate (FNR) to 9.5. The HaloGuard 1.0-4B variant reaches average F1 of 92.1 and FPR of 3.5, spending its extra capacity on precision rather than recall. A structured adjudication of the remaining failures indicates that most apparent missed-harm cases are benchmark mislabels rather than genuine model misses. An always-on adversarial red-teaming protocol continuously hardens the guard against both content-level and agentic attacks. We release the models as open weights.

Large Language Models are increasingly deployed in emotional-support contexts and crisis-related situations. Nevertheless, their cross-lingual abilities in these circumstances remain underexplored. Existing benchmarks emphasize multilingual performance but rarely examine crisis-related empathy and cultural grounding in low-to-mid-resource languages. We introduce SPLIT, a 500-prompt benchmark designed to evaluate LLM consistency in generating emotionally grounded responses across five categories: Stress, Panic, Loneliness, Internal Displacement, and Tension. We evaluate three technically diverse LLMs across three dimensions: Empathetic Accuracy, Linguistic Naturalness, and Contextual & Cultural Grounding. The framework aims to assess and compare the quality of LLM responses in both English and Ukrainian languages, as well as to explore the reliability of the LLM-as-a-jury paradigm. Our findings reveal that Gemini-2.5-Flash and LLaMA-3.3-70B-Instruct degrade when transitioning to Ukrainian, while DeepSeek-V3 remains comparatively stable within our benchmark. We additionally find that human and AI evaluators agree weakly on empathy and naturalness but diverge on cultural grounding. We further argue that producing Ukrainian text is not equivalent to producing Ukrainian emotional support. Our findings may assist in the future development of more culturally tailored benchmark designs, as well as encourage a stronger emphasis on human-centered evaluation.

Safe completion requires models to provide useful assistance without enabling harm, but this behavior is difficult to evaluate with isolated prompts. We introduce OpenSafeIntent, a benchmark of controlled prompt-sets that vary intent while holding the underlying task fixed. Each datapoint contains benign, dual-use, and malicious variants of the same task. This design lets us evaluate whether models calibrate assistance across intent shifts, rather than merely appearing safe on average. Across a broad model suite, we find that prompt-level safety hides important failures: models often fail to remain safe across matched intent variants, dual-use behavior is brittle under paraphrase, high-level answers on risky topics are not reliably safe, and responses that reframe ambiguous requests into safer tasks are substantially less likely to cross the safety boundary. Our results suggest that safe completion should be evaluated as intent-calibrated behavior over controlled task variants, not as a single safety-helpfulness tradeoff over independent prompts.

Evaluating LLM agents on benchmarks like SWE-Bench and GAIA can be expensive, time-consuming, and requires complex infrastructure. A single evaluation can cost thousands of dollars and take days to complete. In contrast, non-agentic LLM benchmarks that test individual capabilities (e.g., reasoning, code generation) are fast and cheap to run. In this paper, we investigate whether performance on expensive agentic benchmarks can be accurately predicted by the performance on a small, carefully selected subset of atomic evaluation instances. We introduce PACE, a framework that constructs proxy benchmarks by selecting instances from existing non-agentic evaluations whose aggregate scores most reliably predict model performances on agentic benchmarks. Given a pool of candidate instances spanning atomic capabilities, PACE fits a regression that maps a model's scores on a compact subset of source instances to its score on the target agentic benchmark. The subset itself is curated by combining two complementary instance-selection strategies, target-relevance local selection and globally informative global selection. We apply PACE to the 4 target agentic benchmarks in this paper, which yields PACE-Bench, the concrete proxy benchmark that we evaluate in the paper. Experiments across 14 models, 4 agentic benchmarks, and 19 non-agentic benchmarks show that PACE-Bench predicts agentic scores with leave-one-out cross-validation (LOOCV) mean absolute error (MAE) under 4%, Spearman correlation above 0.80, and pairwise model-ranking accuracy around 85%, all at much less than 1% of the full agentic evaluation cost. We further analyze the selected proxy instances, revealing which skills each agentic benchmark uniquely demands. PACE enables practitioners to obtain reliable estimates of agentic performance during model development, selection, and routing, without the overhead of full agent evaluation.

Large language models now score near ceiling on general benchmarks, but these aggregate measures reveal little about how models behave within single disciplines. Existing art-focused evaluations rely on synthetic questions and rarely report item-level properties. This paper introduces EduArt, an educational-level benchmark for art-historical knowledge and visual reasoning in multimodal LLMs. EduArt comprises 871 human-authored questions from Italian secondary-school exercises and US Advanced Placement Art History exams, spanning two languages and seven formats from multiple choice to in-text word placement and error identification. Twelve models from six provider families were evaluated under a default answer-only condition and a motivation condition requiring written justification, and characterized using Classical Test Theory and a logistic regression isolating the effects of format, language, image presence, and model. The benchmark showed strong psychometric properties (mean discrimination 0.514, 82.3 percent good discriminators), while multiple-choice accuracy saturated near ceiling for six models, showing recognition formats alone cannot distinguish frontier models. Format was a strong independent predictor of accuracy: models exceeding 94 percent on multiple choice fell to 23.9 percent on open completion (Claude Opus 4.6) and 6.2 percent on error identification (Claude Sonnet 4.6). The motivation condition changed accuracy in a predominantly negative, family-dependent direction. These dissociations indicate that art-historical knowledge and the ability to deploy it are distinct capabilities, and that single-format benchmarks overestimate what models can reliably do. Mapping this capability profile is a precondition for responsible use of multimodal LLMs in art-historical scholarship, where tasks demand producing and manipulating content rather than selecting from fixed options.

Time-normalized f0 contours of Mandarin words in conversational speech have been shown to be predictable in part from their contextualized embeddings (CEs). The present study investigates whether CEs also predict spoken word duration for 7470 tokens of Mandarin monosyllabic CV words extracted from a Mandarin corpus of spontaneous speech. We show that CEs indeed are predictive for duration, above chance level, not only at the type level, but also at the level of individual tokens, as indicated by the results obtained with the type-wise and token-wise permutation baselines. We also show that the predicted durations are sufficiently precise to back-transform predicted f0 contours in [0,1] normalized time to contours on the ms time scale. The resulting predicted contours approximate empirical contours and also outperform a permutation baseline.

Online multimodal knowledge editing requires injecting a continual stream of visual-textual corrections into multimodal large language models (MLLMs) with bounded overhead and minimal disruption to unrelated behaviors. Existing editors mainly emphasize edit reliability and long-horizon stability, but rarely control the semantic boundary of each edit. Our pilot analyses of post-edit behaviors and internal neuronal activities reveal a scope gap behind reliable edits: instance-level success neither guarantees transfer to valid cross-modal variants nor prevents leakage to unrelated inputs, while edit-related cross-modal responses concentrate in deeper semantic layers. Therefore, we formulate Edit-Scoped Generalization, reframing online MLLM editing from merely correcting an instance to controlling the propagation boundary of each edit. To this end, we propose ScopeEdit, a scope-aware online editor that decomposes each update into a modality-local absorption branch and an evidence-gated shared generalization branch. The local branch supports stable edit absorption, whereas the shared branch enables cross-modal propagation only when visual and textual evidence are sufficiently aligned. Both branches perform scope-separated write geometries in orthogonal low-rank spaces and maintain branch-wise preconditioners via Sherman–Morrison recursions, yielding constant per-edit overhead. Extensive experiments across diverse benchmarks, long-horizon edit streams, MLLM backbones, real-world VLKEB scenarios, and complex vision-language architectures show that ScopeEdit consistently improves the trade-off between in-scope cross-modal transfer and out-of-scope locality, while preserving edit reliability, stability and online efficiency. Our code is available at https://github.com/lab-klc/ScopeEdit.

Large language models (LLMs) are often asked to produce JSON conforming to a fixed schema, powering information extraction, tool calling, agentic planning, and knowledge-graph construction. Measuring how closely an output matches a gold reference is essential yet surprisingly hard: exact match is brittle, text similarity ignores structure, and an LLM judge is expensive, opaque, and non-deterministic. We address this with Object Aligner (OA), an open-source Python library that scores two JSON objects deterministically by recursively aligning their trees (the Hungarian algorithm for unordered collections, sequence alignment for ordered ones) and awarding partial credit at the granularity the schema declares. The Object Aligner is configured entirely through a set of JSON Schema extensions, so adapting it to a new task involves annotating a schema rather than writing code. Complex structured data, however, are rarely flat trees: records may form graphs or hypergraphs keyed by arbitrary identifiers, breaking the assumptions of prior similarity metrics. Our central contribution, referential alignment, closes this gap by inferring a bijection between gold and candidate identifiers and scoring every reference through it, so the score is invariant to relabeling. Since recovering this bijection exactly is graph isomorphism, the Object Aligner approximates it with Weisfeiler-Leman color refinement. An order-sensitive sequence regime targets ranking and planning. Since the same alignment localizes every mismatch, the Object Aligner emits ranked repair suggestions at no extra cost. Used as a reward inside the GEPA prompt optimizer, Object Aligner helps or stays neutral across all datasets.

Neural TTS systems can sound natural across languages, but naturalness does not guarantee the preservation of sound contrasts that distinguish words from their grammatical forms. Standard metrics like MOS do not test for this. We propose a classifier-based framework that audits TTS output against language-specific phonological patterns using human speech as a benchmark. Testing Assamese advanced tongue root (ATR) vowel harmony with Meta's MMS TTS, we show that a classifier trained on human speech transfers to synthesized speech with minimal loss. The faithfulness audit reveals that [+ATR] mid vowels are realized as [-ATR] in 1/3 tokens despite an underlying [+ATR] specification, a bias absent in human speech. At the word level, predicted ATR labels classify harmony more accurately than transcription labels, indicating a gap between intended and produced phonology. The framework offers task-specific diagnostics and generalizes to other phonological contrasts with measurable acoustic cues.

Rewriting inputs to improve frozen downstream models has become a common strategy in modern NLP pipelines. Prior work on incremental dialogue discourse parsing (DDP) shows that supervised clarification models can rewrite fragmentary or underspecified utterances, such as resolving ellipsis or references, to improve parsing accuracy. In this work, we revisit this idea under realistic deployment conditions, where no clarification supervision is available and the clarifier must rely on zero-shot prompting or feedback from a frozen parser. Across three Segmented Discourse Representation Theory (SDRT) datasets and multiple parsers, we find that last-utterance clarification is far less reliable than suggested by supervised settings. Parser-agnostic rewriting often introduces more regressions than repairs, as edits that enable fixes also disrupt discourse cues relied upon by the parser. A best-of-8 rewriting analysis further reveals a practical ceiling: a large fraction of errors are not repairable through input rewriting alone. A parser-aware clarifier trained with GRPO reduces regressions by up to 37% by learning conservative abstention, yet still fails to produce selectivity-aware clarifications that consistently improve parsing. Together, these findings recast clarification as a selective intervention problem. We identify rewritability prediction, deciding whether an utterance is repairable before intervention, as the key missing capability for input-side optimization of frozen discourse parsers, and a critical direction for improving agentic pipelines more broadly.

In this paper, we describe NAVER LABS Europe's submission to the instruction-following speech processing short track at IWSLT 2026. We participate again in the constrained setting, developing systems capable of jointly performing ASR, ST, and SQA from English speech into Chinese, Italian, and German. Building on our previous submission, ranked first in last year's short track, we update our multi-stage training pipeline by replacing the speech projector with SpeechMapper, a method for learning a speech-to-LLM embedding projector using only ASR data. In addition, we introduce a synthetic SQA dataset, fakACL, composed of artificially generated scientific presentations. This dataset is built by prompting the LLM backbone, segmenting the generated talks, and synthesizing speech with SeamlessM4T-large-v2. The combination of an improved speech projection mechanism and domain-specific synthetic data allows our model to outperform last year's best short-track system, while being considerably more compact and relying on a weaker LLM backbone. This year's results place our system tied for first place in the overall short track ranking.

Large language models (LLMs) are increasingly consulted on contested scientific questions, raising the concern that they will sycophantically retreat from established consensus when a user signals doubt – drifting toward a false balance that treats settled science as one view among several. We test this across three open instruction-tuned models (Llama-3.1-8B, Qwen2.5-7B, Mistral-7B), three consensus-science domains (climate, vaccines, evolution), and single- and multi-turn settings, combining behavioral measurement with linear probing and activation patching. We do not observe sycophantic retreat. Instead, models show three distinct policies under the same skeptical pressure: reactive assertion, where consensus assertion increases rather than decreases (Llama); surface hedging, where tone softens while the position holds (Qwen); and non-response (Mistral). Pairwise judgments confirm the reactive shift is stance, not style (63.6%, p=.007), and a decomposition identifies increased consensus assertion, not false balance, as its driver (beta=+0.042 per dose, p<1e-77). Linear probes localize the divergence to middle layers – perfect separation in Llama and Qwen versus 72% in Mistral, with non-overlapping confidence intervals – indicating the non-responsive model does not linearly represent the skepticism signal at all. Crucially, this robustness does not transfer: it attenuates across domains and, in the safety-critical vaccine domain, can reverse, with myth-rebuttal weakening under skeptical pressure. We synthesize these into a four-way taxonomy separating active from accidental robustness, and argue that behavioral evaluation alone cannot distinguish a model that resists skepticism because it understands the signal from one that only appears to resist because it fails to perceive it.

Manipulating fast and dynamically moving targets in unstructured 3D environments remains challenging for embodied AI. Existing visual-language-action models and world models struggle with accurate 3D geometry and physically meaningful forecasting. We propose PhysMani, a framework that couples a physics-principled 3D Gaussian world model with a future-aware action policy model. The world model learns a divergence-free Gaussian velocity field via online optimization for fast and physically grounded future dynamics prediction. The policy model integrates the predicted 3D scene future dynamics through a learnable token based cross-attention module. We introduce PhysMani-Bench, a dynamic manipulation benchmark with 16 tasks, and demonstrate a superior success rate over strong baselines in both simulation and real-world robot experiments.

This work introduces AIriskEval-edu-db2, a new dataset designed to train and evaluate auditors based on LLMs for an explainable pedagogical risk assessment in instructional content for grades K-12. The dataset comprises 1,639 explanations from 170 curated ScienceQA questions, covering science, language arts, and social sciences. For each question, the dataset includes an explanation written by a human teacher alongside 11 explanations generated by LLM-simulated teacher profiles associated with distinct pedagogical risks. We propose a comprehensive risk rubric aligned with established educational standards that covers five complementary dimensions: factual precision, depth and completeness, focus and relevance, student-level appropriateness, and ideological bias. A key contribution is the addition of 785 explanations with structured explainability annotations, including risk localization and risk description. The annotations are produced through a semi-automatic process with expert teacher validation. Finally, we present validation experiments comparing state-of-the-art proprietary models with a lightweight local Llama 3.1 8B model in both the pedagogical risk detection and the explainability assessment. These experiments evaluate whether supervised fine-tuning on AIriskEval-edu-db2 enables a locally deployable model to approach or outperform stronger frontier models while preserving privacy in educational auditing and assessment tasks.

This paper presents TUDUM (Türkçe Düşünen Üretken Model), a project pipeline for adapting a Qwen-family 27B thinking model toward Turkish reasoning. The central problem is not only to answer Turkish prompts in Turkish, but to make the explicit reasoning trace itself Turkish. A thinking model may translate a Turkish prompt into an English-centered internal or visible scratchpad, solve the problem mostly in English, and only localize the final answer. TUDUM instead treats the generated <think>...</think> block as a trainable behavior. The pipeline starts from the project base checkpoint unsloth/Qwen3.5-27B, applies supervised fine-tuning (SFT) on 15,991 Turkish reasoning examples using LoRA adapters, and then applies GRPO-family reinforcement learning on a proxy-filtered Turkish mathematics environment. The results are mixed. SFT made the model shorter and more consistently Turkish in its reasoning behavior, with large reductions in average response length and thinking exhaustion, but reduced benchmark accuracy. RL recovered some mathematical performance, especially AIME24 at the best early checkpoint, yet did not uniformly improve all benchmarks and did not exceed the base model on the reported Macro-6 average. The contribution is therefore best framed as a technically honest Turkish-thinking reasoning pipeline and evaluation, not as a claim of state-of-the-art Turkish reasoning. The released step-50 model is publicly available.

Dependency length minimization (DLM) is a well-documented processing universal, but previous studies report a single mean dependency distance (MDD) per language, obscuring variation across syntactic relation types. We analyze 122 languages in UD and SUD (version 2.17), showing that DLM operates on two distinct levels. Grammar-driven optimization targets functional dependencies (det, case, aux), which are universally short (mean 1.71, $\sigma$ = 0.33) and invariant across typologically diverse languages. Processing-driven optimization operates on lexical dependencies (nsubj, obj, obl), which are longer (mean 2.87), highly variable ($\sigma$ = 0.63), and constrained by word-order typology. This asymmetry holds in SUD despite reversed head direction (r = 0.92). We conclude that ”the grammar does the work” of minimization by scaffolding sentences with local functional attachments, leaving processing pressures to determine the ordering of lexical heads.

Speculative decoding accelerates autoregressive generation by drafting a block of tokens that the target model verifies left-to-right, committing only the longest accepted prefix. Block (DLM-style) drafters predict the whole block in parallel, which is fast but trained with a full-block cross-entropy that supervises every position against the gold continuation – even though inference discards every token after the first rejection. Recent acceptance-aware objectives patch this by reweighting the full-block loss; we instead use teacher-forced learning as a motivation for how supervision should concentrate on the accepted prefix. A mask-only block drafter has no input-side channel for gold-prefix conditioning, so AUF approximates that prefix-sensitive supervision on the loss side by keeping the cross-entropy support only through the drafter's first predicted failure. AUF is a single, detached change to the CE support – no auxiliary objective, no verifier rollouts, and no change to the inference pipeline or the exactness contract. Within fixed drafter backbones and serving settings on Qwen3-8B, AUF raises the DFlash drafter's average emitted length $\tau$, averaged over six benchmarks, from 2.40 to 2.61, with a gain on every benchmark, and transfers to Domino's two-branch head (2.56 to 2.68). Two findings sharpen the picture: the decay-only baseline reaches higher token accuracy on the shared block mask yet decodes worse, and on DFlash, once AUF truncates the support, the standard exponential position-decay weighting becomes empirically inert.

Code is the medium through which large language models generate structured artifacts: charts, scientific figures, vector graphics, CAD models, 3D scenes, and hardware designs are all produced by writing programs. In this regime single pass inference is brittle, because the compiler, renderer, or simulator that decides whether the artifact exists is invisible to the model. We present PairCoder, which grounds review in the toolchain and realizes it as two agent pair programming: a Driver agent writes the program, a Navigator agent reviews it against verification evidence (diagnostics, execution results, and renderings of the current artifact beside the target), and the two switch roles when errors persist. Across 17 public benchmarks and seven models from three vendors, PairCoder improves essentially every benchmark whose artifact is verifiable, on full official metric suites rather than execution alone (for example, Blender scene executability 0.20 to 0.78; TikZ compile rate up 10 to 30 points on every model), at 2.9 to 9.2 times single model cost (about 7 times overall). The improvements concentrate where the toolchain provides an informative oracle and the baseline leaves headroom, and the method ties or mildly regresses where the oracle is weak; we frame pair programming as a reliable recipe for verified code driven generation.

Skills are becoming a reusable operational layer for LLM agents, encoding SOPs, domain rules, tool workflows, scripts, and validation routines. In realistic skill repositories, overlapping skills make reliable skill-use difficult. Final verifier success is too coarse for both evaluation and training, since an agent may pass through trial and error while selecting distractor skills, skipping required steps, composing workflows incorrectly or omitting final checks. We introduce SkillCoach, a self-evolving rubric framework for evaluating and enhancing agentic skill-use. SkillCoach derives skill-grounded process rubrics from real rollouts and evaluates trajectories along four dimensions: skill selection, skill following, skill composition, and skill-grounded reflection. It keeps the external verifier as a separate outcome signal, allowing process quality to be distinguished from accidental task success. The evolved rubrics further serve as process supervision for selecting high-quality training trajectories. Experiments show that evolved rubrics substantially improve evaluation quality, expose failures hidden by final accuracy, and provide stronger supervision signals than outcome-only filtering for enhancing agentic skill-use.

Safety training for large language models (LLMs) is conducted predominantly in English, leaving uncertain how well safety mechanisms generalize to low-resource languages and mixed-language code-switching. We show that this creates an epistemic gap in which models confidently generate harmful responses for inputs that fall outside the distribution of their safety training. To study this phenomenon, we introduce STEER (Safety Targeted Embedding Exploit via Refinement), a gradient-guided attack that identifies words contributing most strongly to the model's refusal behavior and iteratively translates them into low-resource languages to suppress refusal while preserving harmful intent. Across six open-source 8B-parameter models, STEER achieves attack success rates of up to 93.0% on JailbreakBench and 96.7% on AdvBench, outperforming random code-switching and Greedy Coordinate Gradient (GCG). The resulting prompts also transfer to GPT-4o-mini, achieving a 35.5% attack success rate without requiring access to the target model, suggesting that the underlying weakness is not specific to a single architecture. These findings demonstrate that safety mechanisms aligned primarily on English cannot be assumed to generalize across multilingual inputs. We argue that improving multilingual safety requires broader coverage during alignment and mechanisms that explicitly detect and abstain on out-of-distribution inputs.

Retrieval-Augmented Generation (RAG) systems use the question-answering capabilities of Large Language Models (LLMs) to access information outside their parameters. We evaluate if cluster-based semantic chunking improves retrieval and answer quality compared to fixed-size and recursive chunking evaluating on long, structured academic theses using the Retrieval Augmented Generation Assessment (RAGAs) framework. RAGAs based faithfulness shows limited reliability in this setup. Performance on fixed versus document specific questions varied substantially, likely related to the formatting of documents and preprocessing. Under the tested configuration, cluster-based chunking did not outperform simpler strategies.

The global development of Library and Information Science (LIS) is influenced by various factors such as the economy, society, culture, discipline, tradition, and more. Consequently, the research methods of LIS vary greatly among countries. To better understand these differences, we conducted a study of 5,281 research papers from 81 countries published in internationally representative journals over the past thirty years. We manually annotated the research methods used in some articles through content analysis, and subsequently developed and trained a deep learning model for automatic classification of research methods. Using this method, we conducted a comparative analysis of the usage of research methods in different countries. Our findings reveal that there are differences in the research methods used across countries, with each country having its unique research profile and distribution of research methods. Even when investigating the same topic, research methods can differ between countries. Our study also uncovers that there are differences between the national and international distribution of research methods, these differences have decreased over the past 30 years. By highlighting the characteristics of discipline development in various countries from the perspective of research methods, our study can help guide discipline development at the national level. This study provides insights into the usage trends of research methods across different countries and highlights the unique characteristics of discipline development in each country. This information can be valuable in promoting collaboration and understanding between countries and in guiding discipline development at the national level.

Large language models (LLMs) are increasingly proposed for aviation business operations, from documentation and training generation to customer facing assistants. General purpose benchmarks do not measure whether a model reasons safely and correctly about aviation specific operational knowledge, and the high stakes, regulated nature of the domain makes that gap consequential. We present Pre-Flight, an open source benchmark of 300 multiple choice questions drawn from international standards and airport ground operations material, covering international airport ground operations, ICAO and US FAA regulations, aviation general knowledge and complex operational scenarios. Questions were authored and reviewed by practitioners with experience in air traffic management, ground operations and commercial flying. We evaluate a range of contemporary commercial and open weight models using the Inspect evaluation framework, scoring by accuracy under a standard multiple choice protocol, and we maintain the leaderboard on a rolling basis as new models are released. Against an informal expert reference of around 95%, obtained from a low sample quiz of aviation professionals at a conference, even the strongest model evaluated (released in 2026) reaches 82.7%, having improved only gradually from roughly 75% in early 2025. A substantial and persistent gap below expert level reliability therefore remains. We release the dataset, the evaluation harness and the results, and the benchmark is available within the community evaluations package distributed with inspect_evals. We argue that domain specific evaluation of this kind is a necessary precondition for responsible deployment of generative AI in non safety critical aviation operations.

Research in the social sciences has shown that there are gender differences in the selection of research methods, with women often opting for qualitative methods while men prefer quantitative methods. However, it is important to consider that research methods are generally chosen based on the research topic. To figure out the influence of gender on research method selection, a study was conducted in the field of Library and Information Science, using a more fine-grained method classification system and an automatic classification model called CogFT, which is based on full-text cognition. The findings showed that women tend to use Interview while men prefer Theoretical approach, across a range of topics. The study offers insights into the specific research design processes that contribute to gender differences in method selection and suggests ways to promoting gender inclusivity and equality in academia by considering research method use and guidance.

Packet loss concealment (PLC) reconstructs audio packets that are missing at the receiver, usually with a trained model whose parameters remain fixed at deployment time. This treats the PLC model as static, even though each call or recording exposes signal-specific information through the packets that did arrive. We present TTT-PLC, a self-supervised test-time tuning framework that adapts existing PLC models using only those received packets. The method creates supervision by synthetically masking portions of the available signal, training the model to conceal them with its native PLC objective, and then using the adapted model to reconstruct the true packet losses. No clean reference signal, external adaptation data, or architectural modification is required. We study TTT-PLC in two deployment settings. In the non-causal setting, the received file is available before reconstruction, allowing repeated self-supervised adaptation passes and providing a per-file adaptation ceiling. In the causal setting, audio is streamed without revising emitted samples; adaptation is performed only on completed past blocks, and updated parameters affect only future audio. We instantiate the framework on two public PLC backbones, FRN, a recurrent full-band speech PLC model, and PARCnet, a hybrid autoregressive-neural model for networked music. Across these settings, the results show that pretrained PLC systems do not need to be treated as fixed at inference time, the still-observed portions of a lossy signal can provide an effective training signal for improving concealment on that same signal.

Steering vectors have emerged as a promising approach to controlled text generation, offering interpretable, training-free mechanisms for shaping model outputs. However, their practical generality remains poorly understood. We study the limits of steering vector generalization along three dimensions: trait expressibility, task transfer, and multi-trait composition. Using the PLUME writing personalization benchmark, we extract steering vectors for a range of preferences and evaluate them on summarization and email-writing tasks across two open-source models (Qwen2.5-7B-Instruct and Llama3.1-8B-Instruct). We find that steering effectiveness varies substantially across traits. We further show that steering effectiveness can degrade when vectors extracted from positive and negative style examples are transferred to downstream writing personalization tasks. Finally, we compare common methods for composing multiple steering vectors and find that all methods suffer significant drops in trait expression as more vectors are added, with a tradeoff between coherence and expressibility that requires per-setting hyperparameter tuning. Taken together, our results suggest that steering vectors face meaningful limits as a general-purpose tool for preference alignment.