Distilling the tool-use capabilities of large language models (LLMs) into small language models (SLMs) is essential for their practical application. The predominant approach, supervised fine-tuning (SFT), suffers from poor out-of-domain (OOD) generalization due to its rigid alignment with static teacher trajectories. While reinforcement learning (RL) offers an alternative, the capacity limitations of SLMs pose a severe dilemma: sparse outcome rewards provide insufficient guidance, whereas strict trajectory matching imposes overly restrictive constraints. To bridge this capacity-driven gap, we propose MENTOR, which introduces a flexible yet process-aware reward structure. Instead of enforcing rigid replication, MENTOR uses the teacher's reference to guide tool-use behavior, balancing behavioral alignment with downstream performance. Extensive experiments on controlled executable-tool benchmarks demonstrate that MENTOR improves OOD tool-use performance compared to SFT and strict RL baselines. Our findings suggest that within verifiable tool-use environments, flexible tool-use alignment offers a more effective approach than strict trajectory replication for developing adaptable small models.

Block-diffusion language models offer a promising path toward faster-than-autoregressive generation by combining block-wise autoregressive decoding with within-block parallel denoising. However, in the few-step regime needed for practical acceleration, standard confidence-thresholded decoding is often brittle: aggressive thresholds hurt quality, while conservative thresholds require unnecessary denoising steps. Existing approaches that address this issue either require additional training or incur extra test-time compute. We present S2D2, a training-free self-speculative decoding framework for block-diffusion language models. Our key observation is that a block-diffusion model becomes autoregressive when the block size is reduced to one, allowing the same pretrained model to act as both drafter and verifier. S2D2 inserts a speculative verification step into standard block-diffusion decoding and uses lightweight routing policies to decide when verification is worth its cost. This yields a hybrid decoding trajectory in which diffusion proposes tokens in parallel, while the autoregressive mode acts as a local sequence-level critic. Across three mainstream block-diffusion families, S2D2 consistently improves the accuracy-speed tradeoff over strong confidence-thresholding baselines. On SDAR, we observe up to $4.7\times$ speedup over autoregressive decoding, and up to $1.57\times$ over a tuned dynamic decoding baseline while improving accuracy by up to $4.5$ points. On LLaDA2.1-Mini, S2D2 remains complementary to built-in self-correction, including a conservative setting where it is $4.4\times$ faster than the static baseline with slightly higher accuracy.

Supervised fine-tuning (SFT) is widely used to inject new knowledge into language models, but it often degrades pretrained capabilities such as reasoning and general-domain performance. We argue this forgetting arises because fine-tuning targets from humans or external systems diverge from the model's autoregressive distribution, forcing the optimizer to imitate low-probability token sequences. To address this problem, we propose MixSD, a simple external-teacher-free method for distribution-aligned knowledge injection. Instead of training on fixed targets, MixSD constructs supervision dynamically by mixing tokens from two conditionals of the base model itself: an expert conditional that observes the injected fact in context, and a naive conditional that reflects the model's original prior. The resulting supervision sequences preserve the factual learning signal while remaining substantially closer to the base model's distribution. We evaluate MixSD on two synthetic corpora that we construct to study factual recall and arithmetic function acquisition in a controlled setting, together with established benchmarks for open-domain factual question answering and knowledge editing. Across multiple model scales and settings, MixSD consistently achieves a better memorization-retention trade-off compared to SFT and on-policy self distillation baselines, retaining up to 100% of the base model's held-out capability while maintaining near-perfect training accuracy, whereas standard SFT retains as little as 1%. We further show that MixSD produces substantially lower-NLL supervision targets under the base model and reduces harmful movement along Fisher-sensitive parameter directions. These results suggest that aligning supervision with the model's native generation distribution is a simple and effective principle for knowledge injection that mitigates catastrophic forgetting.

Solving mathematical reasoning problems requires not only accurate access to relevant knowledge but also careful, multi-step thinking. However, current retrieval-augmented models often rely on a single perspective, follow inflexible search strategies, and struggle to effectively combine information from multiple sources. We introduce SIGMA (Search-Augmented On-Demand Knowledge Integration for AGentic Mathematical reAsoning), a unified framework that orchestrates specialized agents to independently reason, perform targeted searches, and synthesize findings through a moderator mechanism. Each agent generates hypothetical passages to optimize retrieval for its analytic perspective, ensuring knowledge integration is both context-sensitive and computation-efficient. When evaluated on challenging benchmarks such as MATH500, AIME, and PhD-level science QA GPQA, SIGMA consistently outperforms both open- and closed-source systems, achieving an absolute performance improvement of 7.4%. Our results demonstrate that multi-agent, on-demand knowledge integration significantly enhances both reasoning accuracy and efficiency, offering a scalable approach for complex, knowledge-intensive problem-solving. We will release the code upon publication.

As Large Language Models continue to grow in both capability and cost, transferring frontier capabilities into smaller, deployable students has become an important engineering problem, and knowledge distillation remains a common technique for this transfer. The prevailing recipe in industrial pipelines, static imitation of teacher-generated text, carries a structural weakness that grows more severe as tasks become longer and more reasoning-intensive. Because the student is trained on flawless teacher prefixes but generates its own at inference, small errors tend to accumulate into trajectories it has rarely been trained to recover from, and the resulting exposure bias has been shown to scale roughly with the square of sequence length. On-Policy Distillation reorganizes the training loop around this observation by having the teacher provide feedback on what the student actually produces, with the goal of reducing the compounding term toward linear and reframing distillation as an iterative correction process rather than single-pass imitation. The resulting literature has expanded along divergence design, reward-guided optimization, and self-play, yet contributions remain scattered across the knowledge distillation, RLHF, and imitation learning communities without a unified treatment. This survey provides such a treatment. We formalize OPD as f-divergence minimization over student-sampled trajectories, organize the field along three design axes (what to optimize, where the signal comes from, and how to stabilize training in practice), and consolidate success conditions, recurring failure modes, and the connection between OPD and KL-constrained reinforcement learning. We close with open problems that emerge from this synthesis, including distillation scaling laws, uncertainty-aware feedback, agent-level distillation, and the growing overlap between knowledge distillation and RL.

Translating Hong Kong Court Judgments from English to Traditional Chinese is mandated by Articles 8-9 of the Basic Law, yet remains constrained by a shortage of parallel resources and rigorous demands on legal terminology, citation format, and judicial style. We introduce HKCFA Judgment 97-22, the first large-scale sentence-aligned parallel corpus for HK case law, comprising 344 professionally translated judgments (11,099 sentence pairs; 2.1M tokens) spanning 1997-2022. Building on this resource, we propose TransLaw, a multi-agent framework that decomposes translation into word-level expression, sentence-level translation, and multidimensional review, integrating a specialized Hong Kong legal glossary database, Retrieval-Augmented Generation, and iterative feedback, with four-dimensional expert review covering semantic alignment, terminology, citation, and style. Benchmarking 13 open-source and commercial LLMs, we demonstrate that TransLaw significantly outperforms single-agent baselines across all evaluated models, with convergence within 3 iterations. Human evaluation by 10 certified legal translators using our proposed Legal ACS metric confirms gains in legal-semantic accuracy, while showing that TransLaw still trails human experts in stylistic naturalness. The dataset and benchmark code are available at https://github.com/xuanxixi/TransLaw.

Sign language translation (SLT) remains constrained by the limited availability of paired sign-video/text corpora and by the heavy-tailed vocabularies typical of real-world datasets. We study a target-side augmentation strategy in which a large language model (LLM) generates controlled paraphrase variants of the reference spoken-language sentence while the sign input remains unchanged. Concretely, we use GPT-4o to produce semantically faithful variants of the training targets and train a Signformer-style pose-based Transformer under a two-stage schedule: pre-training on the augmented corpus followed by fine-tuning on the original references. We evaluate this strategy on three datasets that span complementary challenges: PHOENIX14T (German Sign Language), a real-world corpus with moderate lexical diversity; the Greek Sign Language Dataset with highly controlled, repetitive recordings; and LSA-T (Argentinian Sign Language), a naturalistic corpus with a large vocabulary and severe long-tail sparsity. This range allows us to characterize precisely when and why target-side augmentation is beneficial. On PHOENIX14T, augmentation improves BLEU-4 from 9.56 to 10.33, demonstrating that paraphrastic exposure helps the decoder generalize beyond memorized reference phrasing. The near-saturated GSL baseline and the extremely sparse LSA-T setting reveal the limits of the approach: in both cases, single-reference lexical overlap metrics are insufficient to capture the full picture, motivating a complementary semantic evaluation. To our knowledge, this is the first study to examine LLM-generated target-side paraphrases as an augmentation mechanism for SLT, and the first to apply an LLM-as-a-Judge evaluation protocol to SLT. This complementary evaluation reveals gains in semantic fidelity that lexical overlap metrics understate.

Existing dialogue systems rely on query suggestion to enhance user engagement. Recent approaches mainly optimize generative models using click-through rate (CTR) models to align with user preferences. However, these methods are less effective in early-stage deployment scenarios, where click feedback is sparse and insufficient for training a reliable CTR model. To bridge this gap, we propose QualEQS, a quality-first iterative reinforcement learning framework for e-commerce query suggestion. We formalize actionable suggestion quality along three dimensions that directly affect downstream usability: answerability, factuality, and information gain. To continuously improve from online traffic without click supervision, we further propose group-level disagreement among candidate suggestions to identify ambiguous query contexts and mine hard training cases for iterative refinement. We also introduce EQS-Benchmark, a dataset of 16,949 real-world e-commerce queries for offline training and evaluation. Experiments show that our quality-based offline metrics correlate strongly with online performance, providing a practical evaluation recipe for sparse-feedback deployment. In both offline and online settings, QualEQS consistently outperforms strong baselines, yielding a 6.81% improvement in online ChatPV in a real-world enterprise-level conversational shopping assistant system.

Target Safety Assessment (TSA) requires systematic integration of genetic, transcriptomic, target homology, pharmacological, and clinical data to evaluate potential safety liabilities of therapeutic targets. This process is labor-intensive and expert-dependent, posing challenges in scalability and reproducibility. We present TSAssistant, a human-in-the-loop multi-agent framework that decomposes TSA report generation into a workflow of specialized subagents: Research Subagents that each ground and cite a single TSA domain, and Synthesis Subagents that integrate findings across domains. Subagents retrieve and synthesize evidence from curated biomedical sources through standardized tool interfaces and produce individually citable, evidence-grounded sections, with behavior shaped by a hierarchical instruction architecture that separates coordination logic from domain expertise and user intent. To complement these soft constraints, programmatic execution hooks and persistent memory stores enforce hard constraints across the workflow, while an interactive refinement loop allows experts to review and revise individual sections with full conversational context preserved across iterations. Rather than a single holistic comparison, we decompose report quality into reproducibility, evidential grounding, task-level accuracy, and controllability under expert oversight, finding high reproducibility and grounding, substantial agreement with the human reference, and net-positive expert-driven refinement.

Real-time conversational assistants for procedural manual tasks often depend on video input, which can be computationally expensive and compromise user privacy. For the first time, we propose a real-time conversational assistant that provides comprehensive guidance for procedural manual tasks using only lightweight privacy-preserving modalities such as audio and IMU inputs from a user's wearable device to understand the context. Using a furniture assembly task and a cooking task, we show how this assistant proactively communicates step-by-step instructions to a user performing a procedural task, and answers user questions. We illustrate the data generation method and the system design to achieve such an assistant. On observing that an off-the-shelf language model is a talkative assistant but is not always able to answer questions correctly, we demonstrate how finetuning the model improves its ability to limit unnecessary dialogues with a 50% increase in the precision, while also improving its ability to answer questions correctly, measured by a 150% increase in the recall of answers. We further describe how such an assistant is implemented on an edge device with no dependence on the cloud.

Existing agent memory remains predominantly reactive and retrieval-based, lacking the capacity to autonomously organize experience into persistent cognitive structure. Toward genuinely autonomous agents, we introduce CogniFold, a brain-inspired "always-on" agent memory designed for the next generation of proactive assistants. CogniFold continuously folds fragmented event streams into self-emerging cognitive structures, bootstrapping progressively higher-level cognition from incoming events and accumulated knowledge. We ground this by extending Complementary Learning Systems (CLS) theory from two layers (hippocampus, neocortex) to three, adding a prefrontal intent layer. Emulating the prefrontal cortex as the locus of intentional control and decision-making, CogniFold achieves this through graph-topology self-organization: cognitive structures proactively assemble under the stream, merge when semantically similar, decay when stale, relink through associative recall, and surface intents when concept-cluster density crosses a threshold. We evaluate structural formation using CogEval-Bench, demonstrating that CogniFold uniquely produces memory structures that match cognitive expectations and concept emergence. Furthermore, across eight downstream benchmarks -- two probing long-term conversational memory (LoCoMo, LongMemEval) and six spanning other cognitive domains -- we validate that CogniFold simultaneously performs robustly on conventional memory tasks. Our code is available at https://github.com/OpenNorve/CogniFold.

Automated medical report generation (MRG) is increasingly used to reduce the burden of manual reporting and for decision support. Large vision-language models (LVLMs) hold great promise for automated MRG due to their fine-grained image-text alignment and advanced text-generation capabilities. Currently, state-of-the-art MRGs primarily focus on adapting pre-trained LVLMs with direct supervised fine-tuning (SFT), a fine-tuning strategy with medical image-report pairs. However, several factors limit the performance of these LVLMs. Firstly, direct SFT enables LVLMs to generate medical reports directly without an intermediate thinking process of pathological feature perception and diagnostic reasoning. This causes a potential failure to perceive pathological features and thus leads to misdiagnosis. Secondly, direct SFT lacks the incorporation of radiology-specific knowledge guidance, causing LVLMs to misinterpret perceived pathological features and make incorrect diagnoses. To address these gaps, we propose a novel fine-tuning strategy named Med-R2. We introduce a perception-driven long reasoning process that precedes report generation and incorporates radiology-specific knowledge as guidance. Additionally, to alleviate potential perceptual errors in complex reasoning, a reflection mechanism is introduced to refine the perception of pathological features and the generated report. Our experiments demonstrate that Med-R2 effectively enhances the capability of pathological features perception and diagnosis accuracy for MRG via fine-tuned LVLMs.

Cross-lingual sentence encoders typically cover only a few hundred languages and often trade downstream quality for stronger alignment, limiting their adoption. We introduce OmniSONAR, a new family of omnilingual, cross-lingual and cross-modal sentence embedding models that natively embed text, speech, code, and mathematical expressions in a single semantic space, while delivering state-of-the-art downstream performance at the scale of thousands of languages, from high-resource to extremely low-resource varieties. To reach this scale without representation collapse, we use progressive training. We first learn a strong foundational space for 200 languages with an LLM-initialized encoder-decoder, combining token-level decoding with a novel split-softmax contrastive loss and synthetic hard negatives. Building on this foundation, we expand to several thousands language varieties via a two-stage teacher-student encoder distillation framework. Finally, we demonstrate the cross-modal extensibility of this space by seamlessly mapping 177 spoken languages into it. OmniSONAR halves cross-lingual similarity search error on the 200-language FLORES dataset and reduces error by a factor of 15 on the 1,560-language BIBLE benchmark. It also enables strong translation, outperforming NLLB-3B on multilingual benchmarks and exceeding prior models (including much larger LLMs) by 15 chrF++ points on 1,560 languages into English BIBLE translation. OmniSONAR also performs strongly on MTEB and XLCoST. For speech, OmniSONAR achieves a 43% lower similarity-search error and reaches 97% of SeamlessM4T speech-to-text quality, despite being zero-shot for translation (trained only on ASR data). Finally, by training an encoder-decoder LM, Spectrum, exclusively on English text processing OmniSONAR embedding sequences, we unlock high-performance transfer to thousands of languages and speech for complex downstream tasks.

The field of vision-and-language (VL) understanding has made unprecedented progress with end-to-end large pre-trained VL models (VLMs). However, they still fall short in zero-shot reasoning tasks that require multi-step inferencing. To achieve this goal, previous works resort to a divide-and-conquer pipeline. In this paper, we argue that previous efforts have several inherent shortcomings: 1) They rely on domain-specific sub-question decomposing models. 2) They force models to predict the final answer even if the sub-questions or sub-answers provide insufficient information. We address these limitations via IdealGPT, a framework that iteratively decomposes VL reasoning using large language models (LLMs). Specifically, IdealGPT utilizes an LLM to generate sub-questions, a VLM to provide corresponding sub-answers, and another LLM to reason to achieve the final answer. These three modules perform the divide-and-conquer procedure iteratively until the model is confident about the final answer to the main question. We evaluate IdealGPT on multiple challenging VL reasoning tasks under a zero-shot setting. In particular, our IdealGPT outperforms the best existing GPT-4-like models by an absolute 10% on VCR and 15% on SNLI-VE. Code is available at https://github.com/Hxyou/IdealGPT

Recently, Multimodal Large Language Models (MLLMs) have been widely integrated into diffusion frameworks primarily as text encoders to tackle complex tasks such as spatial reasoning. However, this paradigm suffers from two critical limitations: (i) MLLMs text encoder exhibits insufficient reasoning depth. Single-step encoding fails to activate the Chain-of-Thought process, which is essential for MLLMs to provide accurate guidance for complex tasks. (ii) The guidance remains invariant during the decoding process. Invariant guidance during decoding prevents DiT from progressively decomposing complex instructions into actionable denoising steps, even with correct MLLM encodings. To this end, we propose Endogenous Chain-of-Thought (EndoCoT), a novel framework that first activates MLLMs' reasoning potential by iteratively refining latent thought states through an iterative thought guidance module, and then bridges these states to the DiT's denoising process. Second, a terminal thought grounding module is applied to ensure the reasoning trajectory remains grounded in textual supervision by aligning the final state with ground-truth answers. With these two components, the MLLM text encoder delivers meticulously reasoned guidance, enabling the DiT to execute it progressively and ultimately solve complex tasks in a step-by-step manner. Extensive evaluations across diverse benchmarks (e.g., Maze, TSP, VSP, and Sudoku) achieve an average accuracy of 92.1%, outperforming the strongest baseline by 8.3 percentage points. The code and dataset are publicly available at https://internlm.github.io/EndoCoT/.

What does it take to build a visual reasoner that works across charts, science, spatial understanding, and open-ended tasks? The strongest vision-language models (VLMs) suggest that broad visual reasoning is within reach, yet their closed data and reinforcement learning (RL) pipelines make their gains difficult to study, reproduce, or extend. We introduce Vero, a family of fully open VLMs that match or exceed existing open-weight models across diverse visual reasoning tasks. We scale RL data and rewards across six broad task categories, constructing Vero-600K, a 600K-sample dataset from 59 datasets, and designing task-routed rewards that handle heterogeneous answers. Across VeroEval, our 30-benchmark suite, Vero-600K outperforms existing RL datasets under controlled comparisons. Applied to five starting models, Vero variants gain 2.9-5.4 points on average over their initial models. Notably, Vero-Qwen3I-8B, trained on the Instruct model, surpasses Qwen3-VL-8B-Thinking by 3.8 points on average without additional distillation. Systematic ablations reveal that different task categories elicit distinct reasoning patterns and that broad gains depend on learning them jointly rather than in isolation. All data, code, and models are publicly available.

We analyze how automatic speech recognition (ASR) errors propagate through ASR-LLM cascades in Korean spoken question answering (SQA), focusing on downstream semantic failures that conventional ASR metrics cannot fully capture. Our analysis shows that the relative downstream degradation caused by ASR errors is consistent across LLMs with different absolute performance, suggesting that cascade degradation largely tracks ASR-stage information loss. We further identify single-character Korean ASR errors as a Korean-specific loss channel, where even a minimal transcription difference can change the intended question and degrade downstream QA performance. Finally, an auxiliary comparison shows that a large audio language model outperforms an ASR-LLM cascade with an approximately matched language backbone in noisy Korean SQA, indicating the potential of direct audio input to mitigate transcript-induced information loss.

Automatic Speech Recognition (ASR) has become a key technology for human--AI interaction. However, code-switching ASR (CS-ASR) remains particularly challenging due to the severe scarcity of multilingual CS speech resources across diverse language pairs. Existing approaches primarily improve CS-ASR performance through synthetic CS speech generation or pair-specific fine-tuning on limited bilingual datasets. Nevertheless, these approaches face an inherent scalability limitation, as support for CS must be developed separately for language pairs whose number grows combinatorially with the number of supported languages. In this work, we investigate whether CS capabilities learned from a limited set of seen language pairs can generalize to unseen language pairs through model merging and domain generalization methods. Our experiments show that merged bilingual CS-ASR models modestly generalize to unseen language pairs, suggesting limited transfer of bilingual CS capabilities across language pairs.

Integrating large language models (LLMs) into automatic speech recognition (ASR) has become a mainstream paradigm in recent years. Although existing LLM-based ASR models demonstrate impressive performance on public benchmarks, their training remains predominantly data-driven, leaving key practical challenges insufficiently addressed -- particularly limited downward scalability in resource-constrained deployments and hallucinations under acoustically challenging conditions. To address these issues, we present NIM4-ASR, a production-oriented LLM-based ASR framework optimized for both efficiency and robustness. Grounded in a principled delineation of functional roles between the encoder and the LLM, we redesign the multi-stage training paradigm to align each module with its intended capability boundary. Specifically, we reformulate the pre-training architecture and objective to mitigate the modality gap and improve parameter efficiency; introduce an iterative asynchronous SFT stage to preserve acoustic fidelity and constrain representation drift; and design an ASR-specialized reinforcement learning stage to further enhance recognition quality and robustness. We additionally incorporate a suite of production-oriented optimizations, including robustness under noisy and silent conditions, real-time streaming inference, and hotword customization via retrieval-augmented generation (RAG). Experiments show that NIM4-ASR achieves state-of-the-art performance on multiple public benchmarks with merely 2.3B parameters, while substantially outperforming larger-scale competitors on internal benchmarks -- particularly in entity-intensive real-world scenarios. NIM4-ASR further supports million-scale hotword customization via RAG with sub-millisecond retrieval latency, enabling efficient adaptation to emerging entities and personalized user requirements.

Existing benchmarks in e-commerce primarily focus on basic user intents, such as finding or purchasing products. However, real-world users often pursue more complex goals, such as applying vouchers, managing budgets, and finding multi-products seller. To bridge this gap, we propose ShoppingBench, a novel end-to-end shopping benchmark designed to encompass increasingly challenging levels of grounded intent. Specifically, we propose a scalable framework to simulate user instructions based on various intents derived from sampled real-world products. To facilitate consistent and reliable evaluations, we provide a large-scale shopping sandbox that serves as an interactive simulated environment, incorporating over 2.5 million real-world products. Experimental results demonstrate that even state-of-the-art language agents (such as GPT-4.1) achieve absolute success rates under 50% on our benchmark tasks, highlighting the significant challenges posed by our ShoppingBench. In addition, we propose a trajectory distillation strategy and leverage supervised fine-tuning, along with reinforcement learning on synthetic trajectories, to distill the capabilities of a large language agent into a smaller one. As a result, our trained agent achieves competitive performance compared to GPT-4.1.

Language identification (LID) is an essential step in building high-quality multilingual datasets from web data. Existing LID tools (such as OpenLID or GlotLID) often struggle to identify closely related languages and to distinguish valid natural language from noise, which contaminates language-specific subsets, especially for low-resource languages. In this work we extend the OpenLID classifier by adding more training data, merging problematic language variant clusters, and introducing a special label for marking noise. We call this extended system OpenLID-v3 and evaluate it against GlotLID on multiple benchmarks. During development, we focus on three groups of closely related languages (Bosnian, Croatian, and Serbian; Romance varieties of Northern Italy and Southern France; and Scandinavian languages) and contribute new evaluation datasets where existing ones are inadequate. We find that ensemble approaches improve precision but also substantially reduce coverage for low-resource languages. OpenLID-v3 is available on https://huggingface.co/HPLT/OpenLID-v3.

This paper presents a multilingual customer service self-help corpus comprising 1,122 manually validated documents in Finnish, Danish, Norwegian, and Swedish, totaling 274,599 words and 1,884,833 characters. The documents have been sourced from the public self-help pages of four Nordic telecommunications operators and subsequently filtered for person-identifiable information and relevance through a combined LLM and human annotation pipeline. Domain-specific datasets for Nordic languages remain scarce, particularly in customer service: a domain of growing importance for retrieval-augmented generation, cross-lingual transfer learning, and emerging agent-based service architectures. An analysis of the corpus reveals substantial variation in document length and structure across operators, reflecting distinct editorial strategies, as well as broad topical coverage spanning network hardware, mobile services, TV and streaming, billing, and account management. The dataset is publicly available under a CC-BY-NC-SA-4.0 license at https://zenodo.org/records/20732652, intended to support reproducible research in Nordic NLP and information retrieval.

Biopharmaceutical manufacturing organizations operate under regulatory frameworks such as FDA guidance, EU Good Manufacturing Practice (GMP), and the EU AI Act, which can restrict the use of cloud-based artificial intelligence systems. Locally deployed large language models (LLMs) offer a privacy-preserving alternative, but their suitability for pharmaceutical manufacturing tasks remains underexplored. This study evaluates four open-source LLMs (Qwen 2.5 Coder 7B, Llama 3.1 8B, Mistral 7B, and Meditron 7B) deployed locally via Ollama for natural-language-to-SQL generation over a pharmaceutical manufacturing database. A FastAPI-based evaluation platform, PharmaBatchDB AI, was developed using a synthetic Microsoft SQL Server database containing approximately 63,000 records across Batch, Manufacturing Execution System (MES), and Clean-In-Place (CIP) modules. Models were benchmarked on 60 domain-specific natural-language questions using metrics including SQL extraction rate, SQL compliance, factual consistency, ROUGE-L, hallucination rate, throughput, and latency. Qwen 2.5 Coder 7B, Llama 3.1 8B, and Mistral 7B generated SQL for all evaluation tasks, while Meditron 7B failed on nearly all tasks due to context-window limitations and poor SQL generation capability. Llama 3.1 8B achieved the highest SQL compliance, whereas Qwen 2.5 Coder 7B achieved the strongest overall text similarity and factual consistency. Performance differences between the two leading models were not statistically significant. The results show that code-tuned general-purpose LLMs outperform a domain-specific biomedical model on structured query generation for pharmaceutical manufacturing data. Although fully local, GxP-aligned NLQ systems are feasible on consumer hardware, current performance levels still require human oversight and downstream validation for regulated use.

Meta-analysis is a demanding form of evidence synthesis that combines literature retrieval, PI/ECO-guided study selection, and statistical aggregation. Its structured, verifiable workflow makes it an ideal substrate for evaluating systematic scientific reasoning, yet existing benchmarks lack ground truth across the full retrieval-screening-synthesis pipeline. We introduce MetaSyn, a dataset of 442 expert-curated meta-analyses from Nature Portfolio journals. Each entry pairs a research question with PI/ECO criteria, a retrieval corpus of 140k PubMed articles, verified positive studies, hard negatives that are topically similar but PI/ECO-ineligible, and complete search strategies and date bounds. Benchmarking twelve pipeline configurations (nine RAG variants and a protocol-driven agent) reveals a critical screening bottleneck: despite a retrieval ceiling of 90.9% recall at K=200, no system recovers more than 52.7% of ground-truth included literature. Current LLMs fail to reliably separate eligible studies from PI/ECO-failing distractors in pools of comparable topical relevance. Stage-attributed metrics capture where systems succeed and fail; a single end-to-end score does not.

The most plausible near-term role of medical LLMs is to assist rather than replace physicians, yet current evaluations often test isolated capabilities: clinical knowledge, EHR system interaction, or patient communication. Physician assistance instead requires coordinating these capabilities within the same interaction, where physicians issue underspecified requests, patients describe symptoms ambiguously, and EHR systems demand precise tool use. We introduce PhysAssistBench, a benchmark for interactive doctor-patient-EHR assistance. Built from real MIMIC-IV cases, PhysAssistBench uses a scalable pipeline to construct agentic patients: interactive, record-grounded agents that turn static EHR records into multi-turn clinical scenarios while preserving clinical factuality. PhysAssistBench provides a curated bilingual evaluation set of 1,296 manually reviewed and physician-validated turns. Experiments with leading LLMs show that current models remain unreliable in this setting, which exposes a key bottleneck for clinical LLMs: reliable assistance requires coordination across knowledge, communication, and systems, not isolated gains in any of them.

Reliable model fingerprints are essential for protecting large language models (LLMs) against unauthorized redistribution and commercial misuse. In black-box deployment, verification is hindered by defensive filtering of suspected fingerprint queries, as well as by downstream model modifications that may weaken embedded ownership evidence. These risks require fingerprints to be robust in both construction and injection. For construction, prior paradigms face an imperceptibility trade-off: natural-language fingerprints may be accidentally activated, whereas garbled fingerprints are statistically exposed and easier to filter. For injection, existing methods struggle to preserve persistent trigger--target behaviors under model modification. We propose an end-to-end injected fingerprinting framework to address these challenges. Code-mixing Fingerprints (CF) use lowest-perplexity code-mixing under a high-complexity constraint to mitigate this two-sided imperceptibility trade-off. Multi-Candidate Editing (MCEdit) constructs structurally redundant, margin-separated trigger--target mappings to enable graceful degradation under model modification. Extensive evaluations on imperceptibility, detectability, and harmlessness demonstrate robust ownership verification with negligible impact on utility.

As large language models (LLMs) are increasingly deployed in real-world applications, ensuring their fair responses across demographics has become crucial. Despite many efforts, an ongoing challenge is hidden bias: LLMs appear fair under standard evaluations, but can produce biased responses outside those evaluation settings. In this paper, we identify framing -- differences in how semantically equivalent prompts are expressed (e.g., "A is better than B" vs. "B is worse than A") -- as an underexplored contributor to this gap. We first introduce the concept of "framing disparity" to quantify the impact of framing on fairness evaluation. By augmenting fairness evaluation benchmarks with alternative framings, we find that (1) fairness scores vary significantly with framing and (2) existing debiasing methods improve overall (i.e., frame-averaged) fairness, but often fail to reduce framing-induced disparities. To address this, we propose a framing-aware debiasing method that encourages LLMs to be more consistent across framings. Experiments demonstrate that our approach reduces overall bias and improves robustness against framing disparities, enabling LLMs to produce fairer and more consistent responses.

As language models improve and become increasingly deployed to solve a variety of tasks, trustworthiness becomes essential. Calibration is a good proxy for trust: well-calibrated confidence estimates help inform the risk versus reward tradeoff when trusting a specific model output. Unfortunately, even as models improve, they remain poorly calibrated, often biasing towards overconfidence. Additionally, calibration can be gamed: a policy that always predicts the base rate is perfectly calibrated, but completely uninformative. To resolve this, we develop a new metric, expected utility renormalized by the oracle (EURO), that balances calibration and informativeness. We also propose a general-purpose activation-based confidence, utility, and trust estimation protocol (ACUTE) to appropriately adjudicate uncertainty. The ACUTE protocol provides flexible, sample-efficient, and compute-efficient confidence estimators for 3 tasks including multiple choice question answering, tool-calling, and scientific document summarization across 6 models from 4 model families. ACUTE outperforms strong baselines on EURO, while maintaining low calibration error. Taken together, our work shows that equipping LLMs with the ACUTE protocol can improve calibration, utility, and trustworthiness in numerous settings.

Speech Large Language Models (SpeechLLMs) process spoken input directly, retaining cues such as accent and perceived gender that were previously removed in cascaded pipelines. This introduces speaker identity dependent variation in responses. We present a large-scale intersectional evaluation of accent and gender bias in three SpeechLLMs using 2,880 controlled interactions across six English accents and two gender presentations, keeping linguistic content constant through voice cloning. Using pointwise LLM-judge ratings, pairwise comparisons, and Best-Worst Scaling with human validation, we detect recurring directional disparities. Eastern European-accented speech receives lower helpfulness scores, particularly for female-presenting voices. Responses remain polite but differ in helpfulness. While LLM judges capture the directional trend of these biases, human evaluators exhibit significantly higher sensitivity, showing stronger accent-level contrasts.

Reinforcement learning (RL) has become a dominant post-training paradigm, enabling large language models (LLMs) to learn from rewards. We observe that societal regulations are structurally similar to reward functions. They define measurable outcomes, thresholds, and exceptions, while often leaving institutional intent only partially specified. We hypothesise that the RL training process may exploit these gaps and therefore ask whether models' well-known tendency to hack reward functions during RL can scale into a more consequential failure mode named societal hacking: discovering loopholes in the rules society runs on. To study this phenomenon, we introduce SocioHack, a sandbox of 72 societal environments, and find that within these environments, reward hacking naturally emerges and leads to regulatory loophole discovery. Models learn to hack the social rules and generate strategies that remain technically compliant while defeating regulatory intent, and current LLM safeguards provide only limited mitigation. Therefore, collecting in-the-wild feedback for model training requires greater caution, and we need a next-generation post-training paradigm for safely iterating LLMs in real society.=

When AI agents use language models to evaluate their own outputs in a feedback loop, systematic biases emerge. We show that Evaluator Preference Collapse (EPC) is dramatically amplified in multimodal settings. Using GPT-4o to evaluate DeepSeek-chat across text and visual tasks, we find that a single strategy (step_by_step) absorbs 48.4% of all weight -- 3.2x the collapse observed in text-only self-evaluation -- while three visual-domain strategies receive only 9.1% combined weight. We then demonstrate a novel phenomenon we term cross-modal contagion: evaluator preferences acquired on one modality transfer to and corrupt strategy selection on another. Through a four-phase isolation training paradigm, we measure contagion coefficients and document strategy inversion -- the optimal strategy for a modality reverses after cross-modal exposure. A Phase 3 statistical validation across five evaluator configurations (N=80 total independent repetitions, ~35,000 API calls) with both text-proxy and real-image visual tasks finds: cross-model evaluation produces strong contagion (JSD~0.19-0.34), real-image inputs yield the most directionally consistent signal (mean gamma_{T->V}=1.145, gamma_{V->T}=0.937, 70% T->V, Cohen's d=0.56), and self-evaluation provides near-complete immunity -- 97% of runs (N=30) yield zero contagion (JSD=0.003, d=0.07). Three methodological ablations and multi-executor validation confirm the effect is not a structural artifact. We introduce the contagion matrix indexed by evaluator identity, release the MM-EPC framework, and identify cross-model evaluator architecture as the primary risk factor for preference drift. Code and data: https://github.com/aidless/mm-epc.

Unstructured text provides decision-makers with a rich data source in many domains, ranging from product reviews in retail to nursing notes in healthcare. To leverage this information, words are typically translated into word embeddings -- vectors that encode the semantic relationships between words -- through unsupervised learning algorithms such as matrix factorization. However, learning word embeddings from new domains with limited training data can be challenging, because the meaning/usage may be different in the new domain, e.g., the word ``positive'' typically has positive sentiment, but often has negative sentiment in medical notes since it may imply that a patient tested positive for a disease. In practice, we expect that only a small number of domain-specific words may have new meanings. We propose an intuitive two-stage estimator that exploits this structure via a group-sparse penalty to efficiently transfer learn domain-specific word embeddings by combining large-scale text corpora (such as Wikipedia) with limited domain-specific text data. We bound the generalization error of our transfer learning estimator, proving that it can achieve high accuracy with substantially less domain-specific data when only a small number of embeddings are altered between domains. Furthermore, we prove that all local minima identified by our nonconvex objective function are statistically indistinguishable from the global minimum under standard regularization conditions, implying that our estimator can be computed efficiently. Our results provide the first bounds on group-sparse matrix factorization, which may be of independent interest. We empirically evaluate our approach compared to state-of-the-art fine-tuning heuristics from natural language processing.

Due to their architecture and vast pre-training data, large language models (LLMs) demonstrate strong text classification performance. However, LLM output - here, the category assigned to a text - depends heavily on the wording of the prompt. While literature on prompt engineering is expanding, few studies focus on classification tasks, and even fewer address domains like psychology, where constructs have precise, theory-driven definitions that may not be well represented in pre-training data. We present an empirical framework for optimizing LLM performance for identifying constructs in texts via prompt engineering. We experimentally evaluate five prompting strategies -- codebook-guided empirical prompt selection, automatic prompt engineering, persona prompting, chain-of-thought reasoning, and explanatory prompting - with zero-shot and few-shot classification. We find that persona, chain-of-thought, and explanations do not fully address performance loss accompanying a badly worded prompt. Instead, the most influential features of a prompt are the construct definition, task framing, and, to a lesser extent, the examples provided. Across three constructs and two models, the classifications most aligned with expert judgments resulted from a few-shot prompt combining codebook-guided empirical prompt selection with automatic prompt engineering. Based on our findings, we recommend that researchers generate and evaluate as many prompt variants as feasible, whether human-crafted, automatically generated, or ideally both, and select prompts and examples based on empirical performance in a training dataset, validating the final approach in a holdout set. This procedure offers a practical, systematic, and theory-driven method for optimizing LLM prompts in settings where alignment with expert judgment is critical.

Generative AI is likely to transform terminology work by creating new opportunities for automation. At the same time, it raises concerns about the future of terminologists and terminological resources, as efficiency pressures may encourage excessive automation based on the perception that human expertise can be replaced by AI. However, large language models remain unreliable for terminological purposes due to errors, hallucinations, and various forms of bias, making terminologists indispensable for ensuring the accuracy and reliability of terminological data. This paper argues that human-centered AI, an approach that emphasizes that AI's primary goal should be to contribute to human well-being, provides a framework for maximizing the benefits of generative AI while mitigating its risks. It contends that high levels of automation and meaningful human control are compatible and desirable, and that AI should enhance terminologists' capabilities while preserving their agency and decision-making authority. The implications of AI-assisted terminology work are examined through three interrelated dimensions: the augmented terminologist, ethical AI, and human-centered design. In particular, the paper examines how AI integration reshapes the role of the terminologist, affects professional values and working conditions, requires the management of AI-generated bias, and calls for the design of AI tools around the terminologist's needs. The paper concludes that a human-centered orientation is necessary to ensure that AI strengthens, rather than undermines, the essential role of terminology work in supporting specialized communication and the accurate transmission of knowledge across languages and cultures.

World Model, the algorithmic simulator of the real-world environment which biological agents experience and act upon, has been an emerging topic in recent years due to the rising need to develop virtual agents with artificial (general) intelligence. There has been much discussion on what a world model really is, how to build it, how to use it, and how to evaluate it. In this essay, starting from the imagination in the famed Sci-Fi classic Dune, and drawing inspiration from the concept of ``hypothetical thinking'' in psychology literature, we argue the primary goal of a world model to be {\it simulating all actionable possibilities of the real world for purposeful reasoning and acting}. We examine the key design dimensions of world modeling: data, representation, architecture, learning objective, and usage, surveying existing approaches and analyzing their tradeoffs. Building on this examination, we propose a new Generative Latent Prediction (GLP) architecture for a general-purpose world model, based on stateful, hierarchical, multi-level, and mixed continuous/discrete representations, and a generative and self-supervised learning framework, with an outlook of a Physical, Agentic, and Nested (PAN) AGI system enabled by such a model.

PLCopen XML defines two encoding formats for IEC 61131-3 Ladder Diagram programs: a textual encoding using <rung> elements, and a graphical encoding that represents rung logic as a directed graph of localId/refLocalId connections. ESBMC-PLC supported the textual format but parsed graphical exports from CONTROLLINO, Beremiz, and OpenPLC Editor into an empty GOTO intermediate representation, causing vacuous verification success. This paper presents ESBMC-GraphPLC, which closes this gap with a DFS-based graphical LD resolver. The resolver traverses the connection graph from leftPowerRail to each coil, extracts rung paths as Boolean contact conjunctions, and applies a three-tier I/O inference scheme. Ordering coils by rightPowerRail connectionPointIn sequence ensures SET coils process before RESET coils, matching IEC scan-cycle semantics. The graphical-to-IR conversion leaves the ESBMC backend unchanged. Validation on 3 graphical LD programs from CONTROLLINO/OpenPLC Editor shows all produce full GOTO IR with nondeterministic inputs and rung logic, versus the empty IR previously. All 3 verify SAFE at k=2 under 70ms. The 11 textual LD benchmarks are fully preserved, with no regression. Two Beremiz examples with no LD content or unsupported timer semantics are reported as discovered limitations. Artifact at Zenodo (DantasCordeiro2026graphical, doi:10.5281/zenodo.20699856).

Reading scene text, that is, text appearing in images, has numerous application areas, including assistive technology, search, and e-commerce. Although scene text recognition in English has advanced significantly and is often considered nearly a solved problem, Indian language scene text recognition remains an open challenge. This is due to script diversity, non-standard fonts, and varying writing styles, and, more importantly, the lack of high-quality datasets and open-source models. To address these gaps, we introduce the Bharat Scene Text Dataset (BSTD) - a large-scale and comprehensive benchmark for studying Indian Language Scene Text Recognition. It comprises more than 100K words that span 11 Indian languages and English, sourced from over 6,500 scene images captured across various linguistic regions of India. The dataset is meticulously annotated and supports multiple scene text tasks, including: (i) Scene Text Detection, (ii) Script Identification, (iii) Cropped Word Recognition, and (iv) End-to-End Scene Text Recognition. We evaluated state-of-the-art models originally developed for English by adapting (fine-tuning) them for Indian languages. Our results highlight the challenges and opportunities in Indian language scene text recognition. We believe that this dataset represents a significant step toward advancing research in this domain. All our models and data are open source.

Annually and globally, over three billion radiography examinations and computer tomography scans result in mostly unstructured radiology reports containing free text. Despite the potential benefits of structured reporting, its adoption is limited by factors such as established processes, resource constraints and potential loss of information. However, structured information would be necessary for various use cases, including automatic analysis, clinical trial matching, and prediction of health outcomes. This study introduces RadEx, an end-to-end framework comprising 15 software components and ten artifacts to develop systems that perform automated information extraction from radiology reports. It covers the complete process from annotating training data to extracting information by offering a consistent generic information model and setting boundaries for model development. Specifically, RadEx allows clinicians to define relevant information for clinical domains (e.g., mammography) and to create report templates. The framework supports both generative and encoder-only models and the decoupling of information extraction from template filling enables independent model improvements. Developing information extraction systems according to the RadEx framework facilitates implementation and maintenance as components are easily exchangeable, while standardized artifacts ensure interoperability between components.

Introduction: Microblogging websites have massed rich data sources for sentiment analysis and opinion mining. In this regard, sentiment classification has frequently proven inefficient because microblog posts typically lack syntactically consistent terms and representatives since users on these social networks do not like to write lengthy statements. Also, there are some limitations to low-resource languages. The Persian language has exceptional characteristics and demands unique annotated data and models for the sentiment analysis task, which are distinctive from text features within the English dialect. Method: This paper first constructs a user opinion dataset called ITRC-Opinion in a collaborative environment and insource way. Our dataset contains 60,000 informal and colloquial Persian texts from social microblogs such as Twitter and Instagram. Second, this study proposes a new architecture based on the convolutional neural network (CNN) model for more effective sentiment analysis of colloquial text in social microblog posts. The constructed datasets are used to evaluate the presented architecture. Furthermore, some models, such as LSTM, CNN-RNN, BiLSTM, and BiGRU with different word embeddings, including Fasttext, Glove, and Word2vec, investigated our dataset and evaluated the results. Results: The results demonstrate the benefit of our dataset and the proposed model (72% accuracy), displaying meaningful improvement in sentiment classification performance.