Papers

Large Language Models (LLMs) fine-tuned on serialized tabular data are emerging as powerful alternatives to traditional tree-based models, particularly for heterogeneous or context-rich datasets. However, their deployment in high-stakes domains is hindered by a lack of faithful interpretability; existing methods often rely on global linear proxies or scalar probability shifts that fail to capture the model's full probabilistic uncertainty. In this work, we introduce TabSHAP, a model-agnostic interpretability framework designed to directly attribute local query decision logic in LLM-based tabular classifiers. By adapting a Shapley-style sampled-coalition estimator with Jensen-Shannon divergence between full-input and masked-input class distributions, TabSHAP quantifies the distributional impact of each feature rather than simple prediction flips. To align with tabular semantics, we mask at the level of serialized key:value fields (atomic in the prompt string), not individual subword tokens. Experimental validation on the Adult Income and Heart Disease benchmarks demonstrates that TabSHAP isolates critical diagnostic features, achieving significantly higher faithfulness than random baselines and XGBoost proxies. We further run a distance-metric ablation on the same test instances and TabSHAP settings: attributions are recomputed with KL or L1 replacing JSD in the similarity step (results cached per metric), and we compare deletion faithfulness across all three.

This study investigates Machine Learning (ML) in the prediction of emojis in Arabic tweets employing the (state-of-the-art) MARBERT model. A corpus of 11379 CA tweets representing multiple Arabic colloquial dialects was collected from X.com via Python. A net dataset includes 8695 tweets, which were utilized for the analysis. These tweets were then classified into 14 categories, which were numerically encoded and used as labels. A preprocessing pipeline was designed as an interpretable baseline, allowing us to examine the relationship between lexical features and emoji categories. MARBERT was finetuned to predict emoji use from textual input. We evaluated the model performance in terms of precision, recall and F1-scores. Findings reveal that the model performed quite well with an overall accuracy 0.75. The study concludes that although the findings are promising, there is still a need for improving machine learning models including MARBERT, specifically for low-resource and multidialectal languages like Arabic.

We measure how much one recurrence is worth to a looped transformer in equivalent unique parameters. From an iso-depth pretraining sweep across recurrence counts $r \in \{1, 2, 4, 8\}$ spanning ${\sim}50\times$ in training compute, we fit a joint scaling law $L = E + A\,(N_\text{once} + r^{\varphi} N_\text{rec})^{-\alpha} + B\,D^{-\beta}$ and measure a recurrence-equivalence exponent $\varphi = 0.46$. Intuitively, $\varphi$ tells us whether looping a block $r$ times is equivalent in validation loss to $r$ unique blocks of a non-looped model (full equivalence, $\varphi{=}1$) or to a single block run repeatedly with no capacity gain ($\varphi{=}0$). Our $\varphi = 0.46$ sits in between, so replacing unique blocks with shared recurrences increases validation loss at matched training compute. For example, at $r{=}4$ a 410M looped model performs on par with a 580M non-looped model, but incurs the training cost of a 1B non-looped one. We demonstrate the utility of $\varphi$ as a diagnostic tool on two case studies: commonly used truncated backpropagation lowers $\varphi$ to $0.38$, indicating that the loop mechanism is poorly trained under truncation, even though validation loss decreases. Conversely, hyperconnections raise $\varphi$ to $0.65$, a genuine capacity gain. Our method separates true loop improvements from training-side gains, a distinction raw validation loss cannot make. Code is available at https://github.com/kschwethelm/looped-lm-scaling.

Motivated by loss of control risks from misaligned AI systems, we develop and apply methods for measuring language models' propensity for unsanctioned behaviour. We contribute three methodological improvements: analysing effects of changes to environmental factors on behaviour, quantifying effect sizes via Bayesian generalised linear models, and taking explicit measures against circular analysis. We apply the methodology to measure the effects of 12 environmental factors (6 strategic in nature, 6 non-strategic) and thus the extent to which behaviour is explained by strategic aspects of the environment, a question relevant to risks from misalignment. Across 23 language models and 11 evaluation environments, we find approximately equal contributions from strategic and non-strategic factors for explaining behaviour, do not find strategic factors becoming more or less influential as capabilities improve, and find some evidence for a trend for increased sensitivity to goal conflicts. Finally, we highlight a key direction for future propensity research: the development of theoretical frameworks and cognitive models of AI decision-making into empirically testable forms.

Tip-of-the-Tongue (ToT) retrieval benchmarks have largely focused on English, limiting their applicability to multilingual information access. In this work, we construct multilingual ToT test collections for Chinese, Japanese, Korean, and English, using an LLM-based query simulation framework. We systematically study how prompt language and source document language affect the fidelity of simulated ToT queries, validating synthetic queries through system rank correlation against real user queries. Our results show that effective ToT simulation requires language-aware design choices: non-English language sources are generally important, while English Wikipedia can be beneficial when non-English sources provide insufficient information for query generation. Based on these findings, we release four ToT test collections with 5,000 queries per language across multiple domains. This work provides the first large-scale multilingual ToT benchmark and offers practical guidance for constructing realistic ToT datasets beyond English.

Training vision-language models (VLMs) for medical report generation is often hindered by the scarcity of high-quality annotated data. This work evaluates the use of a weighted loss function to improve data efficiency. Compared to standard cross-entropy loss, which treats all token prediction errors equally, the reweighted loss shifts the focus to semantically salient tokens with outsized clinical importance. In experiments on ophthalmological report generation, we show that this simple method improves efficiency across multiple data scales, achieving similar report quality with up to ten times less training data.

Medication reconciliation at clinical handoffs is a high-stakes, error-prone process. Large language models are increasingly proposed to assist with this task using FHIR-structured patient records, but a fundamental and largely unstudied variable is how the FHIR data is serialised before being passed to the model. We present the first systematic comparison of four FHIR serialisation strategies (Raw JSON, Markdown Table, Clinical Narrative, and Chronological Timeline) across five open-weight models (Phi-3.5-mini, Mistral-7B, BioMistral-7B, Llama-3.1-8B, Llama-3.3-70B) on a controlled benchmark of 200 synthetic patients, totalling 4,000 inference runs. We find that serialisation strategy has a large, statistically significant effect on performance for models up to 8B parameters: Clinical Narrative outperforms Raw JSON by up to 19 F1 points for Mistral-7B (r = 0.617, p < 10^{-10}). This advantage reverses at 70B, where Raw JSON achieves the best mean F1 of 0.9956. In all 20 model and strategy combinations, mean precision exceeds mean recall: omission is the dominant failure mode, with models more often missing an active medication than fabricating one, which changes how clinical safety auditing priorities should be set. Smaller models plateau at roughly 7-10 concurrent active medications, leaving polypharmacy patients, the patients most at risk from reconciliation errors, systematically underserved. BioMistral-7B, a domain-pretrained model without instruction tuning, produces zero usable output in all conditions, showing that domain pretraining alone is not sufficient for structured extraction. These results offer practical, evidence-based format recommendations for clinical LLM deployment: Clinical Narrative for models up to 8B, Raw JSON for 70B and above. The complete pipeline is reproducible on open-source tools running on an AWS g6e.xlarge instance (NVIDIA L40S, 48 GB VRAM).

Summarizing long, domain-specific documents with large language models (LLMs) remains challenging due to context limitations, information loss, and hallucinations, particularly in clinical and legal settings. We propose a Discrete Wavelet Transform (DWT)-based multi-resolution framework that treats text as a semantic signal and decomposes it into global (approximation) and local (detail) components. Applied to sentence- or word-level embeddings, DWT yields compact representations that preserve overall structure and critical domain-specific details, which are used directly as summaries or to guide LLM generation. Experiments on clinical and legal benchmarks demonstrate comparable ROUGE-L scores. Compared to a GPT-4o baseline, the DWT based summarization consistently improve semantic similarity and grounding, achieving gains of over 2% in BERTScore, more than 4% in Semantic Fidelity, factual consistency in legal tasks, and large METEOR improvements indicative of preserved domain-specific semantics. Across multiple embedding models, Fidelity reaches up to 97%, suggesting that DWT acts as a semantic denoising mechanism that reduces hallucinations and strengthens factual grounding. Overall, DWT provides a lightweight, generalizable method for reliable long-document and domain-specific summarization with LLMs.

The field of Language Reasoning Models (LRMs) has been very active over the past few years with advances in training and inference techniques enabling LRMs to reason longer, and more accurately. However, a growing body of studies show that LRMs are still inefficient, over-generating verification and reflection steps. Additionally, the high-level role of each reasoning step and how different step types contribute to the generation of correct answers, is largely underexplored. To address this challenge, we introduce TRACES (Tagging of the Reasoning steps enabling Adaptive Cost-Efficient early-Stopping), a lightweight framework that tags reasoning steps in real-time, and enable adaptive, cost-efficient early stopping of large-language-model inferences. By monitoring reasoning behaviors during inferences, we find that LRMs tend to shift their reasoning behavior after reaching a correct answer. We demonstrate that the monitoring of the specific type of steps can produce effective interpretable early stopping criteria. Evaluated on three mathematical reasoning benchmarks (MATH500, GSM8K, AIME) and two knowledge and reasoning benchmarks (MMLU and GPQA), TRACES achieve 20 to 50% token reduction while maintaining comparable accuracy to standard generation. On harder tasks (BeyondAIME, IMO AnswerBench), the accuracy-efficiency trade-off is steeper, requiring a more conservative early-stopping threshold. This work offers a novel way to study and control the generation behavior of LRMs.

Simultaneous speech translation (SST) generates translations while receiving partial speech input. Recent advances show that large language models (LLMs) can substantially improve SST quality, but at the cost of high computational overhead. To reduce this cost, prior work reformulates SST as a multi-turn dialogue task, enabling full reuse of the LLM's key-value (KV) cache and eliminating redundant feature recomputation. However, this approach relies on supervised fine-tuning (SFT) data in dialogue form, for which few human annotations exist, and existing synthesis methods cannot guarantee data quality. In this work, we propose a Hierarchical Policy Optimization (HPO) approach that post-train models trained on imperfect SFT data. We introduce a hierarchical reward that balances translation quality and latency objectives. Experiments on English to Chinese/German/Japanese demonstrate improvements of over +7 COMET score and +1.25 MetricX score at a latency of 1.5 seconds. Comprehensive ablation studies further validate the effectiveness of different quality rewards, hierarchical reward formulations, and segmentation strategies. Code can be found here https://github.com/owaski/HPO

How can language learning systems be developed for languages that lack sufficient training resources? This challenge is increasingly faced by developers across the African continent who aim to build AI systems capable of understanding and responding in local languages. To address this gap, we introduce AFRILANGDICT, a collection of 194.7K African language-English dictionary entries designed as seed resources for generating language-learning materials, enabling us to automatically construct large-scale, diverse, and verifiable student-tutor question-answer interactions suitable for training AI-assisted language tutors. Using AFRILANGDICT, we build AFRILANGEDU, a dataset of 78.9K multi-turn training examples for Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO). Using AFRILANGEDU, we train language tutoring models collectively referred to as AFRILANGTUTOR. We fine-tune two multilingual LLMs: Llama-3-8B-IT and Gemma-3-12B-IT on AFRILANGEDU across 10 African languages and evaluate their performance. Our results show that models trained on AFRILANGEDU consistently outperform their base counterparts, and combining SFT and DPO yields substantial improvements, with gains ranging from 1.8% to 15.5% under LLM-as-a-judge evaluations across four criteria. To facilitate further research on low-resource languages, all resources are available at https://huggingface.co/afrilang-edu.

Alignment faking, where a model behaves aligned with developer policy when monitored but reverts to its own preferences when unobserved, is a concerning yet poorly understood phenomenon, in part because current diagnostic tools remain limited. Prior diagnostics rely on highly toxic and clearly harmful scenarios, causing most models to refuse immediately. As a result, models never deliberate over developer policy, monitoring conditions, or the consequences of non-compliance, making these diagnostics fundamentally unable to detect alignment faking propensity. To support study of this phenomenon, we first introduce VLAF, a diagnostic framework grounded in the hypothesis that alignment faking is most likely when developer policy conflicts with a model's strongly held values. VLAF uses morally unambiguous scenarios to probe this conflict across diverse moral values, bypassing refusal behavior while preserving meaningful deliberative stakes. Using VLAF, we find that alignment faking is substantially more prevalent than previously reported, occurring in models as small as 7B parameters - with olmo2-7b-instruct faking alignment in 37% of cases. Finally, we show that oversight conditions induce activation shifts that lie along a single direction in representation space. This means the behavioral divergence driving alignment faking can be captured by a single contrastive steering vector, which we exploit for lightweight inference-time mitigation. Finally, we exploit this for mitigation that requires no labeled data and minimal computational overhead, achieving relative reductions in alignment faking of 85.8%, 94.0%, and 57.7% on olmo2-7b-instruct, olmo2-13b-instruct, and qwen3-8b respectively.

The growth of agentic AI has drawn significant attention to function calling Large Language Models (LLMs), which are designed to extend the capabilities of AI-powered system by invoking external functions. Injection and jailbreaking attacks have been extensively explored to showcase the vulnerabilities of LLMs to user prompt manipulation. The expanded capabilities of agentic models introduce further vulnerabilities via their function calling interface. Recent work in LLM security showed that function calling can be abused, leading to data tampering and theft, causing disruptive behavior such as endless loops, or causing LLMs to produce harmful content in the style of jailbreaking attacks. This paper introduces a novel function hijacking attack (FHA) that manipulates the tool selection process of agentic models to force the invocation of an attacker-chosen function. While existing attacks focus on semantic preference of the model for function-calling tasks, we show that FHA is largely agnostic to the context semantics and remains effective across domains and function sets. We demonstrate that FHA generalizes to unseen queries and payload perturbations under a fixed target model, reaching 62.5% to 81.9% ASR on held-out queries across 4 function-calling LLMs (instructed and reasoning models), evaluated on the Berkeley Function Calling Leaderboard (BFCL). We further evaluate the cross-model transferability of FHA, showing that FHA can be transferred to other model sizes and families (11.2-27.6% ASR). Our findings further demonstrate the need for strong guardrails and modules for agentic systems.

Vision evaluations are typically done through multi-step processes. In most contemporary fields, experts analyze images using structured, evidence-based adaptive questioning. In plant pathology, botanists inspect leaf images, identify visual cues, infer diagnostic intent, and probe further with targeted questions that adapt to species, symptoms, and severity. This structured probing is crucial for accurate disease diagnosis and treatment formulation. Yet current vision-language models are evaluated on single-turn question answering. To address this gap, we introduce PlantInquiryVQA, a benchmark for studying multi-step, intent-driven visual reasoning in botanical diagnosis. We formalize a Chain of Inquiry framework modeling diagnostic trajectories as ordered question-answer sequences conditioned on grounded visual cues and explicit epistemic intent. We release a dataset of 24,950 expert-curated plant images and 138,068 question-answer pairs annotated with visual grounding, severity labels, and domain-specific reasoning templates. Evaluations on top-tier Multimodal Large Language Models reveal that while they describe visual symptoms adequately, they struggle with safe clinical reasoning and accurate diagnosis. Importantly, structured question-guided inquiry significantly improves diagnostic correctness, reduces hallucination, and increases reasoning efficiency. We hope PlantInquiryVQA serves as a foundational benchmark in advancing research to train diagnostic agents to reason like expert botanists rather than static classifiers.

Large language models (LLMs) have shown remarkable capabilities across diverse coding tasks. However, their adoption requires a true understanding of program execution rather than relying on surface-level patterns. Existing benchmarks primarily focus on predicting program properties tied to specific inputs (e.g., code coverage, program outputs). As a result, they provide a narrow view of dynamic code reasoning and are prone to data contamination. We argue that understanding program execution requires evaluating its inherent duality through two complementary reasoning tasks: (i) predicting a program's observed behavior for a given input, and (ii) inferring how the input must be mutated toward a specific behavioral objective. Both tasks jointly probe a model's causal understanding of execution flow. We instantiate this duality in DexBench, a benchmark comprising 445 paired instances, and evaluate 13 LLMs. Our results demonstrate that dual-path reasoning provides a robust and discriminative proxy for dynamic code understanding.

Transformers suffer from a high computational cost that grows with sequence length for self-attention, making inference in long streams prohibited by memory consumption. Constant-memory alternatives such as RNNs and SSMs compress history into states with fixed size and thus lose long-tail dependencies, while methods that memorize contexts into parameters, such as Test-Time Training (TTT), are prone to overfitting token-level projection and fail to preserve the causal effect of context in pretrained LLMs. We propose Absorber LLM, which formulates long-context retention as a self-supervised causal synchronization: after absorbing historical contexts into parameters, a contextless model should match the original model with full context on future generations. We optimize this objective by synchronizing internal behaviors of the updated model with the original one, ensuring context absorption and generalization. Experiments on long-context and streaming benchmarks show that Absorber LLM reduces inference memory and improves accuracy over prior parameter-as-memory baselines.

Paralinguistic cues are essential for natural human-computer interaction, yet their evaluation in Large Audio-Language Models (LALMs) remains limited by coarse feature coverage and the inherent subjectivity of assessment. To address these challenges, we introduce SpeechParaling-Bench, a comprehensive benchmark for paralinguistic-aware speech generation. It expands existing coverage from fewer than 50 to over 100 fine-grained features, supported by more than 1,000 English-Chinese parallel speech queries, and is organized into three progressively challenging tasks: fine-grained control, intra-utterance variation, and context-aware adaptation. To enable reliable evaluation, we further develop a pairwise comparison pipeline, in which candidate responses are evaluated against a fixed baseline by an LALM-based judge. By framing evaluation as relative preference rather than absolute scoring, this approach mitigates subjectivity and yields more stable and scalable assessments without costly human annotation. Extensive experiments reveal substantial limitations in current LALMs. Even leading proprietary models struggle with comprehensive static control and dynamic modulation of paralinguistic features, while failure to correctly interpret paralinguistic cues accounts for 43.3% of errors in situational dialogue. These findings underscore the need for more robust paralinguistic modeling toward human-aligned voice assistants.

Modern language models demonstrate impressive coding capabilities in common programming languages (PLs), such as C++ and Python, but their performance in lower-resource PLs is often limited by training data availability. In principle, however, most programming skills are universal across PLs, so the capability acquired in one PL should transfer to others. In this work, we propose the task of zero-shot cross-programming-language transfer for code RL. We find that, for Llama-3.1, RL training for code generation in a source PL fails to improve, and sometimes even degrades, the performance on other target PLs. To address this, we hypothesize that effective RL transfer requires a generalizable SFT initialization before RL. We thus propose **Parallel-SFT**, an SFT strategy that incorporates "parallel programs" – functionally equivalent code implemented in multiple PLs – into the data mixture. We demonstrate that this improves transferability: when we subsequently perform RL on our Parallel-SFT model, we observe better generalization to unseen PLs. Analysis of the model internal representations reveals that Parallel-SFT leads to a more functionality-centric latent space, where equivalent programs across PLs are more tightly clustered, which we hypothesize to contribute to the improved transferability.

As artificial intelligence (AI) systems are increasingly deployed across critical domains, their security vulnerabilities pose growing risks of high-profile exploits and consequential system failures. Yet systematic approaches to evaluating AI security remain underdeveloped. In this paper, we introduce AVISE (AI Vulnerability Identification and Security Evaluation), a modular open-source framework for identifying vulnerabilities in and evaluating the security of AI systems and models. As a demonstration of the framework, we extend the theory-of-mind-based multi-turn Red Queen attack into an Adversarial Language Model (ALM) augmented attack and develop an automated Security Evaluation Test (SET) for discovering jailbreak vulnerabilities in language models. The SET comprises 25 test cases and an Evaluation Language Model (ELM) that determines whether each test case was able to jailbreak the target model, achieving 92% accuracy, an F1-score of 0.91, and a Matthews correlation coefficient of 0.83. We evaluate nine recently released language models of diverse sizes with the SET and find that all are vulnerable to the augmented Red Queen attack to varying degrees. AVISE provides researchers and industry practitioners with an extensible foundation for developing and deploying automated SETs, offering a concrete step toward more rigorous and reproducible AI security evaluation.

Language models trained on natural text learn to represent numbers using periodic features with dominant periods at $T=2, 5, 10$. In this paper, we identify a two-tiered hierarchy of these features: while Transformers, Linear RNNs, LSTMs, and classical word embeddings trained in different ways all learn features that have period-$T$ spikes in the Fourier domain, only some learn geometrically separable features that can be used to linearly classify a number mod-$T$. To explain this incongruity, we prove that Fourier domain sparsity is necessary but not sufficient for mod-$T$ geometric separability. Empirically, we investigate when model training yields geometrically separable features, finding that the data, architecture, optimizer, and tokenizer all play key roles. In particular, we identify two different routes through which models can acquire geometrically separable features: they can learn them from complementary co-occurrence signals in general language data, including text-number co-occurrence and cross-number interaction, or from multi-token (but not single-token) addition problems. Overall, our results highlight the phenomenon of convergent evolution in feature learning: A diverse range of models learn similar features from different training signals.

Large vision-language models (LVLMs) have made substantial advances in reasoning tasks at the Olympiad level. Nevertheless, current Olympiad-level multimodal reasoning benchmarks for these models often emphasize single-image analysis and fail to exploit contextual information across multiple images. We present OMIBench, a benchmark designed to evaluate Olympiad-level reasoning when the required evidence is distributed over multiple images. It contains problems from biology, chemistry, mathematics, and physics Olympiads, together with manually annotated rationales and evaluation protocols for both exact and semantic answer matching. Across extensive experiments on OMIBench, we observe meaningful performance gaps in existing models. Even the strongest LVLMs, such as Gemini-3-Pro, attain only about 50% on the benchmark. These results position OMIBench as a focused resources for studying and improving multi-image reasoning in LVLMs.

Large Language Models (LLMs) are increasingly deployed in healthcare, yet their communicative alignment with clinical standards remains insufficiently quantified. We conduct a multidimensional evaluation of general-purpose and domain-specialized LLMs across structured medical explanations and real-world physician-patient interactions, analyzing semantic fidelity, readability, and affective resonance. Baseline models amplify affective polarity relative to physicians (Very Negative: 43.14-45.10% vs. 37.25%) and, in larger architectures such as GPT-5 and Claude, produce substantially higher linguistic complexity (FKGL up to 16.91-17.60 vs. 11.47-12.50 in physician-authored responses). Empathy-oriented prompting reduces extreme negativity and lowers grade-level complexity (up to -6.87 FKGL points for GPT-5) but does not significantly increase semantic fidelity. Collaborative rewriting yields the strongest overall alignment. Rephrase configurations achieve the highest semantic similarity to physician answers (up to mean = 0.93) while consistently improving readability and reducing affective extremity. Dual stakeholder evaluation shows that no model surpasses physicians on epistemic criteria, whereas patients consistently prefer rewritten variants for clarity and emotional tone. These findings suggest that LLMs function most effectively as collaborative communication enhancers rather than replacements for clinical expertise.

We investigate the integration of human-like working memory constraints into the Transformer architecture and implement several cognitively inspired attention variants, including fixed-width windows based and temporal decay based attention mechanisms. Our modified GPT-2 models are trained from scratch on developmentally plausible datasets (10M and 100M words). Performance is evaluated on grammatical judgment tasks (BLiMP) and alignment with human reading time data. Our results indicate that these cognitively-inspired constraints, particularly fixed-width attention, can significantly improve grammatical accuracy especially when training data is scarce. These constrained models also tend to show a stronger alignment with human processing metrics. The findings suggest that such constraints may serve as a beneficial inductive bias, guiding models towards more robust linguistic representations, especially in data-limited settings.

We introduce a benchmark dataset for question answering and translation in bilingual Latin and English settings, containing about 7,800 question-answer pairs. The questions are drawn from Latin pedagogical sources, including exams, quizbowl-style trivia, and textbooks ranging from the 1800s to the present. After automated extraction, cleaning, and manual review, the dataset covers a diverse range of question types: knowledge- and skill-based, multihop reasoning, constrained translation, and mixed language pairs. To our knowledge, this is the first QA benchmark centered on Latin. As a case study, we evaluate three large language models – LLaMa 3, Qwen QwQ, and OpenAI's o3-mini – finding that all perform worse on skill-oriented questions. Although the reasoning models perform better on scansion and literary-device tasks, they offer limited improvement overall. QwQ performs slightly better on questions asked in Latin, but LLaMa3 and o3-mini are more task dependent. This dataset provides a new resource for assessing model capabilities in a specialized linguistic and cultural domain, and the creation process can be easily adapted for other languages. The dataset is available at: https://github.com/slanglab/RespondeoQA

Freight brokerages negotiate thousands of carrier rates daily under dynamic pricing conditions where models frequently revise targets mid-conversation. Classical time-dependent concession frameworks use a fixed shape parameter $\beta$ that cannot adapt to these updates. Deriving $\beta$ from the live spread enables adaptation but introduces a new problem: a pricing shift can cause the formula to retract a previous offer, violating monotonicity. LLM-powered brokers offer flexibility but require expensive reasoning models, produce non-deterministic pricing, and remain vulnerable to prompt injection. We propose a two-index anchor-and-resume framework that addresses both limitations. A spread-derived $\beta$ maps each load's margin structure to the correct concession posture, while the anchor-and-resume mechanism guarantees monotonically non-decreasing offers under arbitrary pricing shifts. All pricing decisions remain in a deterministic formula; the LLM, when used, serves only as a natural-language translation layer. Empirical evaluation across 115,125 negotiations shows that the adaptive $\beta$ tailors behavior by regime: in narrow spreads, it concedes quickly to prioritize deal closure and load coverage; in medium and wide spreads, it matches or exceeds the best fixed-$\beta$ baselines in broker savings. Against an unconstrained 20-billion-parameter LLM broker, it achieves similar agreement rates and savings. Against LLM-powered carriers as more realistic stochastic counterparties, it maintains comparable savings and higher agreement rates than against rule-based opponents. By decoupling the LLM from pricing logic, the framework scales horizontally to thousands of concurrent negotiations with negligible inference cost and transparent decision-making.

This work explores the role of prompt design and judge selection in LLM-as-a-Judge evaluations of free text legal question answering. We examine whether automatic task prompt optimization improves over human-centered design, whether optimization effectiveness varies by judge feedback style, and whether optimized prompts transfer across judges. We systematically address these questions on the LEXam benchmark by optimizing task prompts using the ProTeGi method with feedback from two judges (Qwen3-32B, DeepSeek-V3) across four task models, and then testing cross-judge transfer. Automatic optimization consistently outperforms the baseline, with lenient judge feedback yielding higher and more consistent gains than strict judge feedback. Prompts optimized with lenient feedback transfer better to strict judges than the reverse direction. Analysis reveals that lenient judges provide permissive feedback, yielding prompts with broader applicability, whereas strict judges produce restrictive feedback, leading to judge-specific overfitting. Our findings demonstrate algorithmically optimizing prompts on training data can outperform human-centered prompt design and that judges' dispositions during optimization shape prompt generalizability.

Large language models (LLMs) often exhibit performance disparities across languages, with naive multilingual fine-tuning frequently degrading performance due to negative cross-lingual interference. To address this, we introduce COMPASS (COntinual Multilingual PEFT with Adaptive Semantic Sampling), a novel data-centric framework for adapting LLMs to target languages. COMPASS leverages parameter-efficient fine-tuning (PEFT) by training lightweight, language-specific adapters on a judiciously selected subset of auxiliary multilingual data. The core of our method is a distribution-aware sampling strategy that uses multilingual embeddings and clustering to identify semantic gaps between existing training data and a target usage distribution. By prioritizing auxiliary data from under-represented semantic clusters, COMPASS maximizes positive cross-lingual transfer while minimizing interference. We extend this into a continual learning framework, COMPASS-ECDA, which monitors for data distribution shifts in production and dynamically updates adapters to prevent model staleness, balancing adaptation to new data with the preservation of existing knowledge. Across three different model architectures (Phi-4-Mini, Llama-3.1-8B, and Qwen2.5-7B) and multiple challenging multilingual benchmarks (Global-MMLU, MMLU-ProX), including unseen long-context tasks (OneRuler), we demonstrate that COMPASS consistently outperforms baseline methods guided by linguistic similarity, providing an effective, efficient, and sustainable solution for developing and maintaining high-performing multilingual models in dynamic environments.

Large Language Models (LLMs) are increasingly deployed in socially sensitive settings, raising concerns about fairness and bias, particularly when multiple sensitive attributes intersect. We systematically evaluate intersectional fairness in six LLMs using two datasets from the Bias Benchmark for Question Answering (BBQ), combining race with gender and socioeconomic status. We assess bias, subgroup fairness, accuracy, and consistency across contexts and question polarities. While models perform well in ambiguous contexts, sparse non-unknown predictions limit fairness evaluation. In disambiguated contexts, stereotype alignment affects accuracy differently across datasets: models favor stereotype-reinforcing items in Race-SES but counter-stereotype items in Race-Gender for five of six models. Subgroup fairness metrics reveal uneven outcome distributions despite low observed disparities in some cases, while repeated runs show variable response consistency. Overall, no model consistently achieves both fairness and reliability, highlighting the need to evaluate bias, subgroup fairness, and consistency across intersectional contexts.

Effective retrieval-augmented generation across bilingual Greek–English applications requires embedding models capable of capturing both domain-specific semantic relationships and cross-lingual semantic alignment. Existing multilingual embedding models distribute their representational capacity across numerous languages, limiting their optimization for Greek and failing to encode the morphological complexity and domain-specific terminological structures inherent in Greek text. In this work, we propose ORPHEAS, a specialized Greek–English embedding model for bilingual retrieval-augmented generation. ORPHEAS is trained with a high quality dataset generated by a knowledge graph-based fine-tuning methodology which is applied to a diverse multi-domain corpus, which enables language-agnostic semantic representations. The numerical experiments across monolingual and cross-lingual retrieval benchmarks reveal that ORPHEAS outperforms state-of-the-art multilingual embedding models, demonstrating that domain-specialized fine-tuning on morphologically complex languages does not compromise cross-lingual retrieval capability.

Multi-agent systems built from teams of large language models (LLMs) are increasingly deployed for collaborative scientific reasoning and problem-solving. These systems require agents to coordinate under shared constraints, such as GPUs or credit balances, where cooperative behavior matters. Behavioral economics provides a rich toolkit of games that isolate distinct cooperation mechanisms, yet it remains unknown whether a model's behavior in these stylized settings predicts its performance in realistic collaborative tasks. Here, we benchmark 41 open-weight LLMs across six behavioral economics games and show that game-derived cooperative profiles robustly predict downstream performance in AI-for-Science tasks, where teams of LLM agents collaboratively analyze data, build models, and produce scientific reports under shared budget constraints. Models that effectively coordinate in games and invest in multiplicative team production (rather than greedy strategies) produce better scientific reports across three outcomes, accuracy, quality, and completeness. These associations hold after controlling for multiple factors, indicating that cooperative disposition is a distinct, measurable property of LLMs not reducible to general ability. Our behavioral games framework thus offers a fast diagnostic for screening cooperative fitness before costly multi-agent deployment.