Papers

Prompt engineering is effective but labor-intensive, motivating automated optimization methods. Existing methods typically require labeled datasets, which are often unavailable, and produce verbose, repetitive prompts. We introduce PrefPO, a minimal prompt optimization approach inspired by reinforcement learning from human feedback (RLHF). Its preference-based approach reduces the need for labeled data and hyperparameter tuning-only a starting prompt and natural language criteria are needed. PrefPO uses an LLM discriminator to express pairwise preferences over model outputs and provide feedback to an LLM optimizer, iteratively improving performance. We evaluate PrefPO on 9 BIG-Bench Hard (BBH) tasks and IFEval-Hard, a newly-curated, challenging subset of IFEval. PrefPO matches or exceeds SOTA methods, including GEPA, MIPRO, and TextGrad, on 6/9 tasks and performs comparably to TextGrad on IFEval-Hard (82.4% vs 84.5%). Unlike other methods, PrefPO can optimize in both labeled and unlabeled settings. Without labels, PrefPO closely matches its labeled performance on 6/9 tasks, proving effective without ground truth. PrefPO also improves prompt hygiene: we find existing methods produce prompts 14.7x their original length or with 34% repetitive content; PrefPO reduces these issues by 3-5x. Furthermore, both LLM and human judges rate PrefPO's prompts higher than TextGrad's. Finally, we identify prompt hacking in prompt optimizers, where methods game evaluation criteria, and find PrefPO is susceptible at half the rate of TextGrad (37% vs 86%), generating fewer brittle, misaligned prompts.

Large language models (LLMs) are increasingly central to clinician workflows, spanning clinical decision support, medical education, and patient communication. However, current evaluation methods for medical LLMs rely heavily on static, templated benchmarks that fail to capture the complexity and dynamics of real-world clinical practice, creating a dissonance between benchmark performance and clinical utility. To address these limitations, we present MedArena, an interactive evaluation platform that enables clinicians to directly test and compare leading LLMs using their own medical queries. Given a clinician-provided query, MedArena presents responses from two randomly selected models and asks the user to select the preferred response. Out of 1571 preferences collected across 12 LLMs up to November 1, 2025, Gemini 2.0 Flash Thinking, Gemini 2.5 Pro, and GPT-4o were the top three models by Bradley-Terry rating. Only one-third of clinician-submitted questions resembled factual recall tasks (e.g., MedQA), whereas the majority addressed topics such as treatment selection, clinical documentation, or patient communication, with ~20% involving multi-turn conversations. Additionally, clinicians cited depth and detail and clarity of presentation more often than raw factual accuracy when explaining their preferences, highlighting the importance of readability and clinical nuance. We also confirm that the model rankings remain stable even after controlling for style-related factors like response length and formatting. By grounding evaluation in real-world clinical questions and preferences, MedArena offers a scalable platform for measuring and improving the utility and efficacy of medical LLMs.

The recent escalation of the Iran Israel USA conflict in 2026 has triggered widespread global discussions across social media platforms. As people increasingly use these platforms for expressing opinions, analyzing public sentiment from these discussions can provide valuable insights into global public perception. This study aims to analyze global public sentiment regarding the Iran Israel USA conflict by mining user-generated comments from YouTube news channels. The work contributes to public opinion analysis by introducing a privacy preserving framework that combines topic wise sentiment analysis with modern deep learning techniques and Federated Learning. To achieve this, approximately 19,000 YouTube comments were collected from major international news channels and preprocessed to remove noise and normalize text. Sentiment labels were initially generated using the VADER sentiment analyzer and later validated through manual inspection to improve reliability. Latent Dirichlet Allocation (LDA) was applied to identify key discussion topics related to the conflict. Several transformer-based models, including BERT, RoBERTa, XLNet, DistilBERT, ModernBERT, and ELECTRA, were fine tuned for sentiment classification. The best-performing model was further integrated into a federated learning environment to enable distributed training by preserving user data privacy. Additionally, Explainable Artificial Intelligence (XAI) techniques using SHAP were applied to interpret model predictions and identify influential words affecting sentiment classification. Experimental results demonstrate that transformer models perform effectively, and among them, ELECTRA achieved the best performance with 91.32% accuracy. The federated learning also maintained strong performance while preserving privacy, achieving 89.59% accuracy in a two client configuration.

Bibliographic reference extraction and parsing are foundational for citation indexing, linking, and downstream scholarly knowledge-graph construction. However, most established evaluations focus on clean, English, end-of-document bibliographies, and therefore underrepresent the Social Sciences and Humanities (SSH), where citations are frequently multilingual, embedded in footnotes, abbreviated, and shaped by heterogeneous historical conventions. We present a unified benchmark that targets these SSH-realistic conditions across three complementary datasets: CEX (English journal articles spanning multiple disciplines), EXCITE (German/English documents with end-section, footnote-only, and mixed regimes), and LinkedBooks (humanities references with strong stylistic variation and multilinguality). We evaluate three tasks of increasing difficulty – reference extraction, reference parsing, and end-to-end document parsing – under a schema-constrained setup that enables direct comparison between a strong supervised pipeline baseline (GROBID) and contemporary LLMs (DeepSeek-V3.1, Mistral-Small-3.2-24B, Gemma-3-27B-it, and Qwen3-VL (4B-32B variants)). Across datasets, extraction largely saturates beyond a moderate capability threshold, while parsing and end-to-end parsing remain the primary bottlenecks due to structured-output brittleness under noisy layouts. We further show that lightweight LoRA adaptation yields consistent gains – especially on SSH-heavy benchmarks – and that segmentation/pipelining can substantially improve robustness. Finally, we argue for hybrid deployment via routing: leveraging GROBID for well-structured, in-distribution PDFs while escalating multilingual and footnote-heavy documents to task-adapted LLMs.

Large language models (LLMs) are increasingly used to assess moral or ethical statements, yet their judgments may reflect social and linguistic biases. This work presents a controlled, sentence-level study of how grammatical person, number, and gender markers influence LLM moral classifications of fairness. Starting from 550 balanced base sentences from the ETHICS dataset, we generated 26 counterfactual variants per item, systematically varying pronouns and demographic markers to yield 14,850 semantically equivalent sentences. We evaluated six model families (Grok, GPT, LLaMA, Gemma, DeepSeek, and Mistral), and measured fairness judgments and inter-group disparities using Statistical Parity Difference (SPD). Results show statistically significant biases: sentences written in the singular form and third person are more often judged as "fair”, while those in the second person are penalized. Gender markers produce the strongest effects, with non-binary subjects consistently favored and male subjects disfavored. We conjecture that these patterns reflect distributional and alignment biases learned during training, emphasizing the need for targeted fairness interventions in moral LLM applications.

Chemical language models (CLMs) have emerged as promising competitors to popular classical machine learning models for molecular property prediction (MPP) tasks. However, an increasing number of studies have reported inconsistent and contradictory results for the performance of CLMs across various MPP benchmark tasks. In this study, we conduct and analyze hundreds of meticulously controlled experiments to systematically investigate the effects of various factors, such as dataset size, model size, and standardization, on the pre-training and fine-tuning performance of CLMs for MPP. In the absence of well-established scaling laws for encoder-only masked language models, our aim is to provide comprehensive numerical evidence and a deeper understanding of the underlying mechanisms affecting the performance of CLMs for MPP tasks, some of which appear to be entirely overlooked in the literature.

Twitter (now X) has become an important source of social media data for situational awareness during crises. Crisis informatics research has widely used tweets from Twitter to develop and evaluate artificial intelligence (AI) systems for various crisis-relevant tasks, such as extracting locations and estimating damage levels from tweets to support damage assessment. However, recent changes in Twitter's data access policies have made it increasingly difficult to curate real-world tweet datasets related to crises. Moreover, existing curated tweet datasets are limited to past crisis events in specific contexts and are costly to annotate at scale. These limitations constrain the development and evaluation of AI systems used in crisis informatics. To address these limitations, we introduce an agentic workflow for generating crisis-related synthetic tweet datasets. The workflow iteratively generates synthetic tweets conditioned on prespecified target characteristics, evaluates them using predefined compliance checks, and incorporates structured feedback to refine them in subsequent iterations. As a case study, we apply the workflow to generate synthetic tweet datasets relevant to post-earthquake damage assessment. We show that the workflow can generate synthetic tweets that capture their target labels for location and damage level. We further demonstrate that the resulting synthetic tweet datasets can be used to evaluate AI systems on damage assessment tasks like geolocalization and damage level prediction. Our results indicate that the workflow offers a flexible and scalable alternative to real-world tweet data curation, enabling the systematic generation of synthetic social media data across diverse crisis events, societal contexts, and crisis informatics applications.

We introduce Holographic Invariant Storage (HIS), a protocol that assembles known properties of bipolar Vector Symbolic Architectures into a design-time safety contract for LLM context-drift mitigation. The contract provides three closed-form guarantees evaluable before deployment: single-signal recovery fidelity converging to $1/\sqrt{2} \approx 0.707$ (regardless of noise depth or content), continuous-noise robustness $2\Phi(1/\sigma) - 1$, and multi-signal capacity degradation $\approx\sqrt{1/(K+1)}$. These bounds, validated by Monte Carlo simulation ($n = 1{,}000$), enable a systems engineer to budget recovery fidelity and codebook capacity at design time – a property no timer or embedding-distance metric provides. A pilot behavioral experiment (four LLMs, 2B–7B, 720 trials) confirms that safety re-injection improves adherence at the 2B scale; full results are in an appendix.

AI development has a fiction dependency problem. Developers have treated large corpora of modern books, including fiction, as valuable enough to accept substantial cost and legal risk, yet current models still struggle to generate compelling long-form fiction. I term this the "AI-Fiction Paradox," and it is particularly startling because training data strongly shapes model output. This paper offers a theoretically precise account of why fiction resists AI generation by identifying three distinct challenges for current systems. First, fiction depends on what I call narrative causation, a form of plot logic where events must feel both surprising in the moment and retrospectively inevitable. Standard autoregressive generation commits to prose sequentially, creating a practical obstacle to coordinating local surprise with retrospective inevitability across a long narrative. Second, I identify an informational revaluation challenge: fiction repeatedly requires the significance of earlier details to be reinterpreted in light of later developments, a form of long-range reasoning that current systems perform unreliably. Third, drawing on over seven years of collaborative research on sentiment arcs, I argue that fiction that moves us requires multi-scale emotional architecture, the orchestration of sentiment at word, sentence, scene, and arc levels simultaneously. Together, these three challenges help explain both why developers have sought large modern book corpora and why compelling long-form fiction remains so difficult to replicate. The analysis also raises urgent questions about what happens when these challenges are overcome. Fiction concentrates unusually powerful cognitive and emotional patterns for modeling human behavior, and mastery of these patterns by AI systems would represent not just a creative achievement but a potent vehicle for human manipulation at scale.

Full-stream text-to-speech (TTS) for interactive systems must start speaking with minimal delay while remaining controllable as text arrives incrementally. We present VoXtream2, a zero-shot full-stream TTS model with dynamic speaking-rate control that can be updated mid-utterance on the fly. VoXtream2 combines a distribution matching mechanism over duration states with classifier-free guidance across conditioning signals to improve controllability and synthesis quality. Prompt-text masking enables textless audio prompting, removing the need for prompt transcription. Across standard zero-shot benchmarks and a dedicated speaking-rate test set, VoXtream2 achieves competitive objective and subjective results against public baselines despite a smaller model and less training data. In full-stream mode, it runs 4 times faster than real time with 74 ms first-packet latency on a consumer GPU.

Model merging has shown that multitask models can be created by directly combining the parameters of different models that are each specialized on tasks of interest. However, models trained independently on distinct tasks often exhibit interference that degrades the merged model's performance. To solve this problem, we formally define the notion of Cross-Task Interference as the drift in the representation of the merged model relative to its constituent models. Reducing cross-task interference is key to improving merging performance. To address this issue, we propose our method, Resolving Interference (RI), a light-weight adaptation framework which disentangles expert models to be functionally orthogonal to the space of other tasks, thereby reducing cross-task interference. RI does this whilst using only unlabeled auxiliary data as input (i.e., no task-data is needed), allowing it to be applied in data-scarce scenarios. RI consistently improves the performance of state-of-the-art merging methods by up to 3.8% and generalization to unseen domains by up to 2.3%. We also find RI to be robust to the source of auxiliary input while being significantly less sensitive to tuning of merging hyperparameters. Our codebase is available at: https://github.com/pramesh39/resolving_interference

Discrete Diffusion Language Models have emerged as a compelling paradigm for unified multimodal generation, yet their deployment is hindered by high inference latency arising from iterative decoding. Existing acceleration strategies often require expensive re-training or fail to leverage the 2D spatial redundancy inherent in visual data. To address this, we propose Locality-Aware Dynamic Rescue (LADR), a training-free method that expedites inference by exploiting the spatial Markov property of images. LADR prioritizes the recovery of tokens at the ”generation frontier”, regions spatially adjacent to observed pixels, thereby maximizing information gain. Specifically, our method integrates morphological neighbor identification to locate candidate tokens, employs a risk-bounded filtering mechanism to prevent error propagation, and utilizes manifold-consistent inverse scheduling to align the diffusion trajectory with the accelerated mask density. Extensive experiments on four text-to-image generation benchmarks demonstrate that our LADR achieves an approximate 4 x speedup over standard baselines. Remarkably, it maintains or even enhances generative fidelity, particularly in spatial reasoning tasks, offering a state-of-the-art trade-off between efficiency and quality.

Backpropagation dominates modern machine learning, yet it is not the only principled method for optimizing dynamical systems. We propose Kalman World Models (KWM), a class of learned state-space models trained via recursive Bayesian filtering rather than reverse-mode automatic differentiation. Instead of gradient descent updates, we replace parameter learning with Kalman-style gain adaptation. Training becomes online filtering; error signals become innovations. We further extend this framework to transformer-based large language models (LLMs), where internal activations are treated as latent dynamical states corrected via innovation terms. This yields a gradient-free training and adaptation paradigm grounded in control theory. We derive stability conditions, analyze computational complexity, and provide empirical results on sequence modeling tasks demonstrating competitive performance with improved robustness and continual adaptation properties.

Instruction Tuning (IT) has been proven to be an effective approach to unlock the powerful capabilities of large language models (LLMs). Recent studies indicate that excessive IT data can degrade LLMs performance, while carefully selecting a small subset of high-quality IT data can significantly enhance their capabilities. Therefore, identifying the most efficient subset data from the IT dataset to effectively develop either specific or general abilities in LLMs has become a critical challenge. To address this, we propose a novel and efficient framework called NAIT. NAIT evaluates the impact of IT data on LLMs performance by analyzing the similarity of neuron activation patterns between the IT dataset and the target domain capability. Specifically, NAIT captures neuron activation patterns from in-domain datasets of target domain capabilities to construct reusable and transferable neuron activation features. It then evaluates and selects optimal samples based on the similarity between candidate samples and the expected activation features of the target capabilities. Experimental results show that training on the 10% Alpaca-GPT4 IT data subset selected by NAIT consistently outperforms methods that rely on external advanced models or uncertainty-based features across various tasks. Our findings also reveal the transferability of neuron activation features across different capabilities of LLMs. In particular, IT data with more logical reasoning and programmatic features possesses strong general transferability, enabling models to develop stronger capabilities across multiple tasks, while a stable core subset of data is sufficient to consistently activate fundamental model capabilities and universally improve performance across diverse tasks.

Large Language Models (LLMs) increasingly serve as autonomous reasoning agents in decision support, scientific problem-solving, and multi-agent coordination systems. However, deploying LLM agents in consequential applications requires assurance that their reasoning remains stable under semantically equivalent input variations, a property we term semantic invariance. Standard benchmark evaluations, which assess accuracy on fixed, canonical problem formulations, fail to capture this critical reliability dimension. To address this shortcoming, in this paper we present a metamorphic testing framework for systematically assessing the robustness of LLM reasoning agents, applying eight semantic-preserving transformations (identity, paraphrase, fact reordering, expansion, contraction, academic context, business context, and contrastive formulation) across seven foundation models spanning four distinct architectural families: Hermes (70B, 405B), Qwen3 (30B-A3B, 235B-A22B), DeepSeek-R1, and gpt-oss (20B, 120B). Our evaluation encompasses 19 multi-step reasoning problems across eight scientific domains. The results reveal that model scale does not predict robustness: the smaller Qwen3-30B-A3B achieves the highest stability (79.6% invariant responses, semantic similarity 0.91), while larger models exhibit greater fragility.

The ability to provide trustworthy maternal health information using phone-based chatbots can have a significant impact, particularly in low-resource settings where users have low health literacy and limited access to care. However, deploying such systems is technically challenging: user queries are short, underspecified, and code-mixed across languages, answers require regional context-specific grounding, and partial or missing symptom context makes safe routing decisions difficult. We present a chatbot for maternal health in India developed through a partnership between academic researchers, a health tech company, a public health nonprofit, and a hospital. The system combines (1) stage-aware triage, routing high-risk queries to expert templates, (2) hybrid retrieval over curated maternal/newborn guidelines, and (3) evidence-conditioned generation from an LLM. Our core contribution is an evaluation workflow for high-stakes deployment under limited expert supervision. Targeting both component-level and end-to-end testing, we introduce: (i) a labeled triage benchmark (N=150) achieving 86.7% emergency recall, explicitly reporting the missed-emergency vs. over-escalation trade-off; (ii) a synthetic multi-evidence retrieval benchmark (N=100) with chunk-level evidence labels; (iii) LLM-as-judge comparison on real queries (N=781) using clinician-codesigned criteria; and (iv) expert validation. Our findings show that trustworthy medical assistants in multilingual, noisy settings require defense-in-depth design paired with multi-method evaluation, rather than any single model and evaluation method choice.

As corporate responsibility increasingly incorporates environmental, social, and governance (ESG) criteria, ESG reporting is becoming a legal requirement in many regions and a key channel for documenting sustainability practices and assessing firms' long-term and ethical performance. However, the length and complexity of ESG disclosures make them difficult to interpret and automate the analysis reliably. To support scalable and trustworthy analysis, this paper introduces ESG-Bench, a benchmark dataset for ESG report understanding and hallucination mitigation in large language models (LLMs). ESG-Bench contains human-annotated question-answer (QA) pairs grounded in real-world ESG report contexts, with fine-grained labels indicating whether model outputs are factually supported or hallucinated. Framing ESG report analysis as a QA task with verifiability constraints enables systematic evaluation of LLMs' ability to extract and reason over ESG content and provides a new use case: mitigating hallucinations in socially sensitive, compliance-critical settings. We design task-specific Chain-of-Thought (CoT) prompting strategies and fine-tune multiple state-of-the-art LLMs on ESG-Bench using CoT-annotated rationales. Our experiments show that these CoT-based methods substantially outperform standard prompting and direct fine-tuning in reducing hallucinations, and that the gains transfer to existing QA benchmarks beyond the ESG domain.

Large Language Models (LLMs) have demonstrated remarkable capability in machine translation on high-resource language pairs, yet their performance on low-resource translation still lags behind. Existing post-training methods rely heavily on high-quality parallel data, which are often scarce or unavailable for low-resource languages. In this paper, we introduce WALAR, a reinforcement training method using only monolingual text to elevate LLMs' translation capabilities on massive low-resource languages while retaining their performance on high-resource languages. Our key insight is based on the observation of failure modes (or "holes") in existing source-based multilingual quality estimation (QE) models. Reinforcement learning (RL) using these QE models tends to amplify such holes, resulting in poorer multilingual LLMs. We develop techniques including word alignment and language alignment to mitigate such holes in WALAR's reward for RL training. We continually trained an LLM supporting translation of 101 languages using WALAR. The experiments show that our new model outperforms LLaMAX, one of the strongest open-source multilingual LLMs by a large margin on 1400 language directions on Flores-101 dataset.

Supervised Semantic Differential (SSD) is a mixed quantitative-interpretive method that models how text meaning varies with continuous individual-difference variables by estimating a semantic gradient in an embedding space and interpreting its poles through clustering and text retrieval. SSD applies PCA before regression, but currently no systematic method exists for choosing the number of retained components, introducing avoidable researcher degrees of freedom in the analysis pipeline. We propose a PCA sweep procedure that treats dimensionality selection as a joint criterion over representation capacity, gradient interpretability, and stability across nearby values of K. We illustrate the method on a corpus of short posts about artificial intelligence written by Prolific participants who also completed Admiration and Rivalry narcissism scales. The sweep yields a stable, interpretable Admiration-related gradient contrasting optimistic, collaborative framings of AI with distrustful and derisive discourse, while no robust alignment emerges for Rivalry. We also show that a counterfactual using a high-PCA dimension solution heuristic produces diffuse, weakly structured clusters instead, reinforcing the value of the sweep-based choice of K. The case study shows how the PCA sweep constrains researcher degrees of freedom while preserving SSD's interpretive aims, supporting transparent and psychologically meaningful analyses of connotative meaning.

Training capable software engineering (SWE) agents demands large-scale, executable, and verifiable environments that provide dynamic feedback loops for iterative code editing, test execution, and solution refinement. However, existing open-source datasets remain limited in scale and repository diversity, while industrial solutions are opaque with unreleased infrastructure, creating a prohibitive barrier for most academic research groups. We present OpenSWE, the largest fully transparent framework for SWE agent training in Python, comprising 45,320 executable Docker environments spanning over 12.8k repositories, with all Dockerfiles, evaluation scripts, and infrastructure fully open-sourced for reproducibility. OpenSWE is built through a multi-agent synthesis pipeline deployed across a 64-node distributed cluster, automating repository exploration, Dockerfile construction, evaluation script generation, and iterative test analysis. Beyond scale, we propose a quality-centric filtering pipeline that characterizes the inherent difficulty of each environment, filtering out instances that are either unsolvable or insufficiently challenging and retaining only those that maximize learning efficiency. With $891K spent on environment construction and an additional $576K on trajectory sampling and difficulty-aware curation, the entire project represents a total investment of approximately $1.47 million, yielding about 13,000 curated trajectories from roughly 9,000 quality guaranteed environments. Extensive experiments validate OpenSWE's effectiveness: OpenSWE-32B and OpenSWE-72B achieve 62.4% and 66.0% on SWE-bench Verified, establishing SOTA among Qwen2.5 series. Moreover, SWE-focused training yields substantial out-of-domain improvements, including up to 12 points on mathematical reasoning and 5 points on science benchmarks, without degrading factual recall.

Long conversations with an AI agent create a simple problem for one user: the history is useful, but carrying it verbatim is expensive. We study personalized agent memory: one user's conversation history with an agent, distilled into a compact retrieval layer for later search. Each exchange is compressed into a compound object with four fields (exchange_core, specific_context, thematic room_assignments, and regex-extracted files_touched). The searchable distilled text averages 38 tokens per exchange. Applied to 4,182 conversations (14,340 exchanges) from 6 software engineering projects, the method reduces average exchange length from 371 to 38 tokens, yielding 11x compression. We evaluate whether personalized recall survives that compression using 201 recall-oriented queries, 107 configurations spanning 5 pure and 5 cross-layer search modes, and 5 LLM graders (214,519 consensus-graded query-result pairs). The best pure distilled configuration reaches 96% of the best verbatim MRR (0.717 vs 0.745). Results are mechanism-dependent. All 20 vector search configurations remain non-significant after Bonferroni correction, while all 20 BM25 configurations degrade significantly (effect sizes |d|=0.031-0.756). The best cross-layer setup slightly exceeds the best pure verbatim baseline (MRR 0.759). Structured distillation compresses single-user agent memory without uniformly sacrificing retrieval quality. At 1/11 the context cost, thousands of exchanges fit within a single prompt while the verbatim source remains available for drill-down. We release the implementation and analysis pipeline as open-source software.

Error Span Detection (ESD) is a crucial subtask in Machine Translation (MT) evaluation, aiming to identify the location and severity of translation errors. While fine-tuning models on human-annotated data improves ESD performance, acquiring such data is expensive and prone to inconsistencies among annotators. To address this, we propose a novel self-evolution framework based on Minimum Bayes Risk (MBR) decoding, named Iterative MBR Distillation for ESD, which eliminates the reliance on human annotations by leveraging an off-the-shelf LLM to generate pseudo-labels. Extensive experiments on the WMT Metrics Shared Task datasets demonstrate that models trained solely on these self-generated pseudo-labels outperform both unadapted base model and supervised baselines trained on human annotations at the system and span levels, while maintaining competitive sentence-level performance.

The widespread adoption of reinforcement learning-based alignment highlights the growing importance of reward models. Various benchmarks have been built to evaluate reward models in various domains and scenarios. However, a significant gap remains in assessing reward models for long-form generation, despite its critical role in real-world applications. To bridge this, we introduce Long-form RewardBench, the first reward modeling testbed specifically designed for long-form generation. Our benchmark encompasses five key subtasks: QA, RAG, Chat, Writing, and Reasoning. We collected instruction and preference data through a meticulously designed multi-stage data collection process, and conducted extensive experiments on 20+ mainstream reward models, including both classifiers and generative models. Our findings reveal that current models still lack long-form reward modeling capabilities. Furthermore, we designed a novel Long-form Needle-in-a-Haystack Test, which revealed a correlation between reward modeling performance and the error's position within a response, as well as the overall response length, with distinct characteristics observed between classification and generative models. Finally, we demonstrate that classifiers exhibit better generalizability compared to generative models trained on the same data. As the first benchmark for long-form reward modeling, this work aims to offer a robust platform for visualizing progress in this crucial area.

Adapting Large Language Models (LLMs) to specialized domains requires high-quality instruction tuning datasets, which are expensive to create through human annotation. Existing data synthesis methods focus on general-purpose tasks and fail to capture domain-specific terminology and reasoning patterns. To address this, we introduce DS$^2$-Instruct, a zero-shot framework that generates domain-specific instruction datasets without human supervision. Our approach first generates task-informed keywords to ensure comprehensive domain coverage. It then creates diverse instructions by pairing these keywords with different cognitive levels from Bloom's Taxonomy. Finally, it uses self-consistency validation to ensure data quality. We apply this framework to generate datasets across seven challenging domains, such as mathematics, finance, and logical reasoning. Comprehensive evaluation demonstrates that models fine-tuned on our generated data achieve substantial improvements over existing data generation methods.

Cyberbullying on social media is inherently multilingual and multi-faceted, where abusive behaviors often overlap across multiple categories. Existing methods are commonly limited by monolingual assumptions or single-task formulations, which restrict their effectiveness in realistic multilingual and multi-label scenarios. In this paper, we propose HMS-BERT, a hybrid multi-task self-training framework for multilingual and multi-label cyberbullying detection. Built upon a pretrained multilingual BERT backbone, HMS-BERT integrates contextual representations with handcrafted linguistic features and jointly optimizes a fine-grained multi-label abuse classification task and a three-class main classification task. To address labeled data scarcity in low-resource languages, an iterative self-training strategy with confidence-based pseudo-labeling is introduced to facilitate cross-lingual knowledge transfer. Experiments on four public datasets demonstrate that HMS-BERT achieves strong performance, attaining a macro F1-score of up to 0.9847 on the multi-label task and an accuracy of 0.6775 on the main classification task. Ablation studies further verify the effectiveness of the proposed components.

Research on developmentally plausible language models has largely focused on English, leaving open questions about multilingual settings. We present a systematic study of compact language models by extending BabyBERTa to English-French scenarios under strictly size-matched data conditions, covering monolingual, bilingual, and cross-lingual settings. Our design contrasts two types of training corpora: (i) child-directed speech (about 2.5M tokens), following BabyBERTa and related work, and (ii) multi-domain corpora (about 10M tokens), extending the BabyLM framework to French. To enable fair evaluation, we also introduce new resources, including French versions of QAMR and QASRL, as well as English and French multi-domain corpora. We evaluate the models on both syntactic and semantic tasks and compare them with models trained on Wikipedia-only data. The results reveal context-dependent effects: training on Wikipedia consistently benefits semantic tasks, whereas child-directed speech improves grammatical judgments in monolingual settings. Bilingual pretraining yields notable gains for textual entailment, with particularly strong improvements for French. Importantly, similar patterns emerge across BabyBERTa, RoBERTa, and LTG-BERT, suggesting consistent trends across architectures.

This paper reports on the development of a leaderboard of Open Large Language Models (LLM) for European Portuguese (PT-PT), and on its associated benchmarks. This leaderboard comes as a way to address a gap in the evaluation of LLM for European Portuguese, which so far had no leaderboard dedicated to this variant of the language. The paper also reports on novel benchmarks, including some that address aspects of performance that so far have not been available in benchmarks for European Portuguese, namely model safeguards and alignment to Portuguese culture. The leaderboard is available at https://huggingface.co/spaces/PORTULAN/portuguese-llm-leaderboard.

Reinforcement Learning with Verifiable Rewards (RLVR) significantly enhances the reasoning capabilities of Large Language Models. When applied to RLVR, Multiple-Choice Questions (MCQs) offer a scalable source of verifiable data but risk inducing reward hacking, where models shortcut reasoning via random guessing or simple elimination. Current approaches often mitigate this by converting MCQs to open-ended formats, thereby discarding the contrastive signal provided by expert-designed distractors. In this work, we systematically investigate the impact of option design on RLVR. Our analysis highlights two primary insights: (1) Mismatches in option counts between training and testing degrade performance. (2) Strong distractors effectively mitigate random guessing, enabling effective RLVR training even with 2-way questions. Motivated by these findings, we propose Iterative Distractor Curation (IDC), a framework that actively constructs high-quality distractors to block elimination shortcuts and promote deep reasoning. Experiments on various benchmarks demonstrate that our method effectively enhances distractor quality and yields significant gains in RLVR training compared to the original data.

Computer Use Agents (CUAs) translate natural-language instructions into Graphical User Interface (GUI) actions such as clicks, keystrokes, and scrolls by relying on a Vision-Language Model (VLM) to interpret screenshots and predict grounded tool calls. However, grounding accuracy varies dramatically across VLMs, while current CUA systems typically route every action to a single fixed model regardless of difficulty. We propose Adaptive VLM Routing (AVR), a framework that inserts a lightweight semantic routing layer between the CUA orchestrator and a pool of VLMs. For each tool call, AVR estimates action difficulty from multimodal embeddings, probes a small VLM to measure confidence, and routes the action to the cheapest model whose predicted accuracy satisfies a target reliability threshold. For warm agents with memory of prior UI interactions, retrieved context further narrows the capability gap between small and large models, allowing many actions to be handled without escalation. We formalize routing as a cost–accuracy trade-off, derive a threshold-based policy for model selection, and evaluate AVR using ScreenSpot-Pro grounding data together with the OpenClaw agent routing benchmark. Across these settings, AVR projects inference cost reductions of up to 78% while staying within 2 percentage points of an all-large-model baseline. When combined with the Visual Confused Deputy guardrail, AVR also escalates high-risk actions directly to the strongest available model, unifying efficiency and safety within a single routing framework. Materials are also provided Model, benchmark, and code: https://github.com/vllm-project/semantic-router.

Reward models (RMs) are critical components of alignment pipelines, yet they exhibit biases toward superficial stylistic cues, preferring better-presented responses over semantically superior ones. Existing debiasing methods typically require retraining or architectural modifications, while direct activation suppression degrades performance due to representation entanglement. We propose SteerRM, the first training-free method for debiasing reward models using Sparse Autoencoder (SAE)-based interventions. SteerRM isolates stylistic effects using contrastive paired responses, identifies bias-related SAE features with a strength-stability criterion, and suppresses them at inference time. Across six reward models on RM-Bench, SteerRM improves Hard-split accuracy by 7.3 points on average while preserving overall performance. Results on a Gemma-based reward model and a controlled non-format bias further suggest generalization across RM architectures and bias types. We further find that format-related features are concentrated in shallow layers and transfer across models, revealing shared architecture-level bias encoding patterns. These results show that SAE-based interventions can mitigate reward-model biases without retraining, providing a practical and interpretable solution for alignment pipelines.