Deployed language models are evaluated in a non-stationary environment: model versions, retrieval layers, safety systems, and real-world inputs all change over time. Static bias benchmarks remain useful, but they do not show how models frame newly emerging events for different prompted audiences. We introduce GPF-LIVENEWS, a streaming evaluation protocol and benchmark snapshot for auditing group-conditioned framing in open-ended LLM outputs. The protocol expands fresh BBC/Reuters news anchors across 42 identity labels and seven prompt families, then evaluates response bundles using semantic-sensitivity and sentiment-disparity signals. In a pilot over 12 monitoring runs and 23 hosted models, Policy/Action prompts produce the strongest semantic movement, while sentiment variation is flatter across dimensions and prompt families. The released artifact includes article metadata, prompt templates, instantiated prompts, model-output metadata, score tables, documentation, and reproduction scripts. We interpret all scores as observed-window audit signals for human review, not as permanent fairness rankings or direct proof of harmful bias.

The success of large language models (LLMs) across diverse NLP tasks has elevated the importance of reasoning chain optimization as a critical step in aligning model behavior with task objectives. Existing reasoning chain tuning methods often rely on black-box heuristics or gradient-free search, which lack interpretability, generalization, and sample efficiency. In this work, we introduce \textbf{Thoughts-as-Planning}, a novel framework that formalizes reasoning chain optimization as a sequential decision-making process over a latent semantic space. We model the LLM as a partially observable environment and learn a latent world model that simulates the effect of reasoning chain edits on downstream outputs. A proximity-preserving embedding space is constructed to encode reasoning chain-response dynamics, enabling planning via gradient descent or reinforcement learning. Our method supports multi-scale abstraction, allowing reasoning chain edits at token, segment, and instruction levels to be integrated into a unified planner. Through extensive experiments on language understanding and generation tasks, we demonstrate that Thoughts-as-Planning outperforms state-of-the-art reasoning chain tuning baselines in efficiency, robustness, and generalization, while offering interpretability through its structured planning trajectory. Our code is available at https://github.com/FastLM/Thoughts-as-Planning.

Large language model (LLM) agents with tool-calling capabilities are increasingly deployed in production systems, yet a fundamental reliability question remains under-explored: does the same agent behave the same way twice? We present a systematic empirical study of behavioral consistency in multi-step tool-calling agents, measuring whether agents select the same tools, in the same order, with the same arguments, across repeated identical invocations. Unlike prior work on consistency in ReAct-style agents(search-only, free-text actions), we study the richer setting of structured tool-calling interfaces with typed parameters and consequential side effects.

Knowledge editing methods such as ROME and MEMIT update factual associations in transformer models by modifying MLP weights. While evaluated mainly by output behavior, their internal mechanism remains underexplored. We investigate whether edits rely on a common mechanism, regardless of which fact is modified. Despite fact-specific weight changes, we argue that ROME and MEMIT target the same subset of weights critical for maintaining edits. To isolate this subset, we train a compact binary mask over the edited weights. The mask reverses 80% of edits on the training set and over 70% on the test set, confirming that diverse edits share a common functional structure. Our analysis reveals that the mask reverses edits by eliminating overattention in later layers. Additionally, we show that injecting the mask during editing drops editing success from 98% to 38%, demonstrating that this mechanism is necessary for edits to succeed. Our finding that edits suppress rather than overwrite knowledge explains why ROME and MEMIT fail to propagate changes to related facts. The identified common functional subspace informs detection and defense against unwanted edits.

Extracting detailed clinical information from free-text medical narratives remains a practical challenge for researchers and healthcare systems. Terminology for immune-mediated and infectious diseases is especially inconsistent across sources, which often limits the ability of general-purpose Natural Language Processing (NLP) systems to capture the relevant biomedical concepts with sufficient granularity. We developed a domain-specific Named Entity Recognition (NER) model tailored to identify disease-related entities occurring in immunology and infectious disease contexts. We assembled and manually annotated a dataset of 371 case reports in collaboration with two clinical specialists, defining twelve entity classes covering immune-mediated and infectious conditions as well as related symptoms and clinical descriptors. We evaluated several modeling strategies, including the MedicalNER architecture with multiple healthcare-specific embeddings, a BERT-based token classification model, and zero-shot NER systems. The strongest performance was obtained with a transformer-based model trained on clinical-domain embeddings, which reached an F1 score of 0.89, consistently outperforming baseline and zero-shot approaches. The combination of specialized embeddings and expert annotation proved particularly valuable for capturing nuanced disease terminology and improving generalization across heterogeneous biomedical text. The prompted LLM baseline achieved substantially lower performance under the same evaluation protocol, reflecting difficulties in producing span-consistent outputs for fine-grained entity boundaries despite detailed prompting. The resulting model provides a structured way to analyze case reports and can support downstream tasks such as cohort identification, disease monitoring, and clinical decision support.

While Large Language Models (LLMs) have demonstrated remarkable capabilities, their reliability is significantly compromised by hallucinations. Existing intrinsic self-correction methods attempt to address this, but often fail due to self-bias, where models struggle to identify errors in their own outputs without external verification. To overcome these limitations, we propose the LDPC-inspired semantic error correction for retrieval-augmented generation (SERC), providing a theoretical framework to interpret and mitigate LLM hallucinations. We reformulate the text generation process as a semantic noisy channel, treating generated responses as noise-corrupted codewords. Inspired by low-density parity-check (LDPC) codes, SERC employs a sparse verification strategy: instead of exhaustively checking all facts, it generates low-density verification queries and validates them against external evidence to efficiently detect and correct errors. We evaluate SERC on LongForm Bio and TruthfulQA benchmarks using Llama-3-8B and Qwen2.5-14B. Experimental results demonstrate that SERC outperforms both intrinsic self-correction methods and strong retrieval-augmented baselines, demonstrating significant gains especially in factual precision (FactScore). Notably, SERC enables small language models (SLMs) to surpass the performance of larger baselines in hallucination reduction and information preservation. Our findings demonstrate that SERC provides a training-free, model-agnostic solution that significantly reduces verification overhead compared to dense methods, achieving an optimal trade-off between cost and fidelity in resource-constrained environments.

Large Language Models (LLMs) extend their capabilities through function-calling (FC), which relies on training data with high quality, diversity, and broad coverage of scenario. However, obtaining and annotating real function-calling data is challenging, while synthetic data from existing pipelines often suffers from unreliable APIs, limited tool scalability, insufficient diversity, and weak quality control. To address these, we present GenesisFunc, an automated pipeline for generating FC training data. Starting from reliable tools in widely used public benchmarks, our GenesisFunc employs a multi-agent framework to support a dialogue generation system that produces conversations spanning diverse scenarios, while maintaining both diversity and quality throughout the process. The accuracy of the data is further reinforced through a multi-stage evaluation system. We fine-tune an 8B LLM on the synthetic dataset and show through extensive experiments that it outperforms similarly sized open-source models in in-domain FC performance and out-of-domain generalization, while reaching FC capabilities comparable to some of the latest API-based models. In addition, our method demonstrates strong potential to scale effectively across downstream tools, underscoring its real-world applicability.

Syllabification describes the task of dividing words into syllables. Due to many rules and exceptions, training an algorithm to perform syllabification with high accuracy remains a challenge. Throughout the last decades, different algorithms have been put forth for Dutch syllabification, yet a comprehensive comparative assessment has not been done. Additionally, deep learning has gained significant popularity within NLP in recent years, yet no modern deep-learning based framework has been developed for Dutch orthographic syllabification. Finally, phonetic and orthographic syllabification algorithms have been examined separately, but not in combination. The aim of the current research was twofold: (a) to examine the performance of existing Dutch syllabification algorithms, and (b) to investigate whether combining phonetic and orthographic information into a single model can increase syllabification performance. To compare the performance of algorithms, four algorithms (Brandt Corstius, Liang, Trogkanis-Elkan (CRF), and a newly conceived deep-learning model) were applied to three different datasets (dictionary words, loanwords, pseudowords). The algorithms show varying performance across datasets, with the data-driven algorithms outperforming a knowledge-based algorithm in all but one condition. The new deep-learning methods developed led to increased performance compared to the best found in the literature (99.65% word accuracy, a 0.14% improvement). An analysis of the words for which adding phonetic information improved syllabification performance indicates that these were words in which the orthographic ambiguity could be resolved by information on pronunciation. Future research could examine other areas where phonetic information can benefit orthographic processing. In addition, the newly developed deep learning frameworks can be applied to other languages than Dutch.

Automatic speech recognition (ASR) has the potential to substantially reduce manual annotation effort in child speech research by generating automatic transcriptions. However, obtaining reliably high-quality ASR transcriptions for child speech remains challenging in low-resource languages due to limited child-specific pre-trained models and highly diverse noise conditions. This study investigates the effectiveness of state-of-the-art ASR models on child speech through two research questions, by evaluating nine ASR models from three model families (Whisper, Parakeet, and Wav2Vec2) on two Dutch child speech datasets, JASMIN and DART. Research question 1 examines the performance of ASR-models applied to child speech. The fine-tuned Whisper-medium model achieves the best overall performance, with a WER of 5.54% on JASMIN and 70.37% on DART, showing that the noisy DART data are clearly more challenging. Research question 2 examines to what extent it is possible to select a subset for which reliable orthographic transcriptions can be obtained automatically, without the need for manual verification. We use an utterance-level selection method that compares ASR output with the original read prompt to identify correctly pronounced recordings. Using the proposed selection method, 42.0% [for JASMIN] and 18.1% [for DART] of the utterances can be automatically identified as correctly pronounced with high confidence, resulting in very low error rates on an utterance level (precisions of 98.3% and higher) and reducing the need for manual verification.

Topic modeling is a branch of Natural Language Processing (NLP) that aims to organize large collections of texts into coherent groups according to word co-occurrence patterns, with Latent Dirichlet Allocation (LDA) remaining one of the most widely used and interpretable probabilistic approaches. Recent advances in NLP, particularly transformer-based language models, offer improved document representations. It is also known that the size of the model (in terms of number of parameters) has a significant impact in the performance of the language models on different pre-defined tasks. In this study, we systematically examine the effect of model size on topic quality by analyzing the performances of seven transformer-based language models (from small models such as MiniLM to large ones such as LLaMA-2) in a BERTopic pipeline on a variety of corpora. Topic quality is evaluated using coherence and divergence metrics following R{ö}der et al. (2015). Our results indicate that model size, ranging from 22 million to 13 billion parameters, has a negligible impact on the quality of the topic, suggesting that smaller models can achieve comparable performance to larger models.

As Large Language Models (LLMs) are increasingly deployed in safety-critical applications, robust content moderation becomes essential. We present a comprehensive evaluation of 14 open-source safety guard models on a curated benchmark of 79,331 samples spanning 8 NIST AI Risk Framework safety categories. Our benchmark aggregates four diverse datasets (HarmBench, StrongREJECT, RealToxicityPrompts, and BeaverTails), filtered to focus exclusively on safety-relevant content (violence, hate speech, harassment, sexual content, suicide/self-harm, profanity, threats, and health misinformation). We find that recall is the critical metric for safety applications, as missing harmful content poses greater risk than false positives. Our evaluation reveals surprising results: Qwen Guard (4B parameters) achieves the highest recall (83.97%) while larger models like Llama Guard (12B) and GPT-OSS Safeguard (20B) exhibit conservative behavior, missing up to 75% of unsafe content. We demonstrate that model size does not correlate with safety detection performance and that general-purpose guard models outperform specialized ones. These findings provide practical guidance for selecting safety guard models in production deployments.

Large Language Models (LLMs) achieve impressive performance across many tasks but remain prone to hallucination, especially in long-form generation where redundant retrieved contexts and lengthy reasoning chains amplify factual errors. Recent studies highlight a critical phenomenon: the closer key information appears to the model outputs, the higher the factual accuracy. However, existing retrieval-augmented language models (RALMs) lack effective mechanisms to ensure this proximity - external evidence is injected into reasoning via multi-turn retrieval, but this cannot ensure key information stays close to the outputs. We propose Micro-Macro Retrieval (M2R), a novel retrieve-while-generate framework to fill this gap. At the macro level, M2R retrieves coarse-grained evidence from external sources; at the micro level, it extracts essential results from a key information repository built during reasoning and reuses them while generating answers. This design directly addresses the key-information-to-output proximity bottleneck, effectively reducing hallucination in long-form tasks. M2R is trained with a curriculum learning-based reinforcement learning strategy using customized rule-based rewards, enabling stable acquisition of retrieval and grounding skills. Extensive experiments across different benchmarks demonstrate the effectiveness of M2R, especially in lengthy-context settings.

Open Arabic large language models split into two classes: sub-1B multilingual models that treat Arabic as an afterthought (Qwen2.5-0.5B, Falcon-H1-0.5B), and 7B-70B Arabic-specialized models that require a server to run (Jais, AceGPT, ALLaM, SILMA). The one published attempt at a sub-2B Arabic-specialized model, Kuwain-1.5B, never released its weights. We present RightNow-Arabic-0.5B-Turbo, a 518M-parameter Arabic-specialized decoder LLM built on Qwen2.5-0.5B. The pipeline adds 27,032 Arabic tokens via mean-subtoken initialization, continues pretraining on 504M Arabic tokens on 8xH100 with FSDP, FlashAttention varlen packing, and Liger fused kernels, then applies supervised fine-tuning on 129,116 Arabic instruction pairs with response-only loss masking, direct preference optimization on 6,750 Arabic preference pairs, and weight soup merging across three checkpoints. On three lm-evaluation-harness Arabic benchmarks (COPA-ar, Arabic HellaSwag, ArabicMMLU) the merged model reaches 35.9% mean accuracy, beats every same-class open model, ties Falcon-H1-1.5B on COPA-ar (58.4%) at one-third the size, and recovers 67% of SILMA-9B's mean at 1/18 the parameters. The edge build quantizes to 398 MB (q4_k_m) and delivers 635 tokens/s at batch size 1 on a single H100 via llama.cpp. All code (5,555 lines across 25 scripts), weights (bf16, int8, and four GGUF quantizations), and benchmark scripts are released at https://huggingface.co/RightNowAI/RightNow-Arabic-0.5B-Turbo.

In modern LLMs, linguistic features function not as stylistic artifacts but as probes of probability mass, allocated under training alignment objectives. Language models trained with contemporary pipelines exhibit severe reshaping of linguistic features, leading to extreme language re-distribution. While previous stylometric analyses explored linguistic differences between AI-generated and human texts, we focus on the reshaping plaguing the LLM training pipeline itself. We analyze 17 models (410M-100B+ parameters) across 24 linguistically-motivated probes, documenting that instruction-tuned systems systematically collapse language entropy along discourse and structural dimensions (mean amplification: 1,949-16,853%, peaks: 5,181-209,675%), while selectively suppressing complex punctuation to 3.2-23.2% of baseline frequencies. These effects do not worsen under RLHF, as divergence patterns are statistically indistinguishable (p > 0.25) across matched base and instruction-tuned model pairs. Weak intervention (lambda=1.0) exacerbates collapse by 240%, while strong control (lambda=5.0) achieves 40.5% improvement and outperforms frontier models by 96.7-98.2% despite 200-1000x scale disadvantage. Additionally, lambda=5.0 delivers 15% higher distinct-4, 27% higher vocabulary diversity, and 78% lower repetition than moderate regularization, establishing that alignment requires sufficient control strength, not merely distributional smoothing. Our findings underscore how modern LLMs reallocate stylistic probability mass, despite RLHF and scale. More broadly, our work reveals a structural limitation of current alignment pipelines: preference optimization reshapes language distributions invisible to standard quality metrics yet detectable through distributional probes, with implications for AI detection, training data contamination, and long-term linguistic evolution.

Large language models (LLMs) frequently encode factual and reasoning knowledge in their internal representations that is not faithfully reflected in their surface-level outputs -- a phenomenon known as \emph{latent knowledge}. Existing approaches to eliciting latent knowledge, such as Contrastive Consistency Search (CCS), rely on contrastive activation patterns and struggle with complex multi-step reasoning tasks, while mechanistic interpretability tools have primarily been used to \emph{understand} model behavior rather than to \emph{extract} hidden knowledge. We present \textbf{MechELK}, a unified three-stage framework that bridges mechanistic interpretability and latent knowledge elicitation. MechELK operates through: (1) \textbf{Locate} -- using Sparse Autoencoder (SAE) feature analysis and activation patching to identify knowledge-bearing representations; (2) \textbf{Verify} -- employing causal probing to distinguish genuine latent knowledge from spurious correlations; and (3) \textbf{Elicit} -- applying representation engineering to surface hidden knowledge without modifying model weights. Evaluated on TruthfulQA, a curated Deceptive Alignment benchmark, and the Quirky LM dataset, MechELK achieves an average elicitation accuracy of 84.7\%, outperforming CCS by 6.2\% and direct linear probing by 9.1\%. Crucially, MechELK successfully identifies latent knowledge in 78.3\% of cases where the model's surface output is incorrect or evasive, demonstrating its utility for AI safety applications including deceptive alignment detection.

Constructing artificial lexicons that are pronounceable, typologically plausible, and semantically structured remains an open challenge in computational linguistics. Existing conlang generators either lack formal phonotactic guarantees or delegate generation to opaque, non-reproducible LLM-based pipelines. We propose a modular framework that samples phoneme inventories from PHOIBLE, generates word forms under interchangeable phonological grammars (deterministic, OT, and MaxEnt), and assigns meanings via a Swadesh--Leipzig--Jakarta ontology with explicit form--meaning alignment. Evaluation on character $n$-gram perplexity, log-likelihood, and KL divergence against PHOIBLE across lexicon sizes of 100-5,000 forms shows that probabilistic grammars consistently outperform deterministic and random baselines on both phonotactic coherence and typological realism.

As the influence of LLMs expands, it is imperative to gain insight into their decisions. One way to do that is to develop probes that detect the presence or absence of a broad set of concepts within the embeddings computed in an LLM - which is what we might say a model is "thinking" about. Such probes should be low-cost and easily applicable to any LLM, so that monitoring for many concepts is possible during normal operation. In this paper, we take the first steps towards developing the capability of creating many such probes by defining and executing examples of the key tasks needed: first, the careful delineation of a concept through the creation of a dataset with the concept both present and then absent. Then, the training and testing of a set of linear probes to detect the concept on any layer of an LLM, including an exploration of the complexity of the probe needed. Finally, we show that such probes can track concepts across larger contexts. This is done with four separate concepts and three different LLMs. When this process is scaled to many more concepts, it will create the ability to easily monitor new models.

Many imaging problems require computing spatial transformations induced by spatially varying intensity, feature, or density fields. Canonical examples include distortion correction, deformable image registration, atlas-based segmentation, and deformation-driven image analysis. These tasks can be formulated as geometric mapping problems in which the transformation is constrained to preserve local structure, control boundary behavior, or regulate angular distortion. Such formulations typically lead to variational models, diffusion processes, or elliptic partial differential equations. However, repeatedly solving high-resolution systems becomes computationally expensive when the underlying parameter fields vary across instances. In this work, we propose a resolution-free neural surrogate for geometric parameterization and mapping problems. Given a spatially varying parameter field $p:Ω\to\mathbb{R}^m$ and query locations $\{x_i\}_{i=1}^N\subsetΩ$, the model predicts mapped locations $\{u(x_i)\}_{i=1}^N$ on arbitrary structured or unstructured point sets. To avoid dependence on a fixed grid, we use a multi-resolution geometric encoding strategy that conditions the network on coordinate-augmented samples of the parameter field. The model is trained without labeled solution data by enforcing geometry-aware constraints derived from variational energies, diffusion-based density equalization, and quasi-conformal theory. Experimental results on quasi-conformal mapping and density-equalizing mapping problems are presented to demonstrate the effectiveness of our proposed method.

Proactive Recommender Systems (PRSs) aim to guide user preference shift toward target items by generating paths of intermediate recommendations. Reinforcement learning (RL) provides a principled framework for optimizing such sequential decision tasks, as path rewards can naturally capture both short-term acceptance and long-term guidance effectiveness. However, naively applying policy gradients to PRS results in deficient gradient estimation. We identify two deficiencies: (1) path-level rewards decompose into step-level rewards with positive mean, creating a length-dependent bias that causes gradients to favor path extension over meaningful exploration; (2) weighting each step by the entire path-level reward ignores the decomposition structure, leading to high gradient variance. To rectify these two deficiencies, we propose an effective RL framework ProRL with two novel mechanisms for proactive recommendation. First, Stepwise Reward Centering subtracts expected rewards to neutralize length-dependent bias, ensuring that path extension yields zero expected gradient signal. Second, Position-Specific Advantage Estimation leverages the reward decomposition structure to compute step-dependent baselines, reducing gradient variance. Together, these mechanisms yield policy gradients that precisely target path quality. Our experiments on three real-world datasets demonstrate that ProRL significantly outperforms state-of-the-art PRSs. Our code is available at https://github.com/hongruhou89/ProRL.

A long-lived LLM agent, such as OpenClaw, earns its value by acting on a user's preferences and constraints across sessions, not just the current request. Yet today's agents keep what a user volunteers but rarely ask for what stays unspoken, leaving a proactivity gap in long-lived LLM agents: an agent cannot act on a preference it never obtained. As users delegate more of their affairs to agents, the impact of this gap grows. We isolate one concrete, controllable slice of this gap as Ask-to-Remember (ATR): the agent decides whether to ask now for a reusable user preference that the current task does not need but a later session with the same user will. ATR is hard even to evaluate: the right question is underdetermined and its payoff deferred to tasks that may never arise. ATRBench, to the best of our knowledge the first ATR benchmark, makes it measurable by fixing each user's preferences as hidden ground truth, so success demands asking, not recall. Across eight frontier LLM agents, defaults fall at least 62 points below an oracle handed the relevant preference, and prompting closes little of it. Diagnostics identify acquisition as the bottleneck. ATRBench surfaces this proactivity gap in current agents and offers a diagnostic testbed for closing it.

Large language model (LLM)-based agents have shown strong capabilities in using external tools to solve complex tasks. However, existing evaluations often overlook the temporal dimension of tool use, especially the impact of tool response latency, and are usually limited to single-task settings. In real-world applications, multiple tasks often need to be executed concurrently, and overall efficiency depends on whether an agent can use idle time while waiting for tool responses. We refer to this capability as asynchronous tool calling. To evaluate it, we propose AsyncTool, a benchmark for assessing LLM-based agents in interactive multi-task tool-use environments with delayed tool feedback. AsyncTool presents multiple heterogeneous tasks simultaneously and simulates realistic tool response latency during execution. Using a hybrid data evolution strategy, we construct a diverse asynchronous multitasking dataset that covers multiple scenarios and tool-use patterns. We evaluate models at the step, sub-task, and task levels, and introduce efficiency-oriented metrics to measure task coordination and completion efficiency. Extensive experiments show that delayed tool feedback poses substantial challenges to current agents and leads to clear performance degradation. Models that better coordinate task switching, dependency tracking, and state maintenance achieve stronger performance on AsyncTool. Our analysis identifies key failure modes of current tool-using agents and provides practical insights for designing future systems with stronger temporal reasoning and coordination capabilities.

Generative models, including diffusion models, are increasingly used as foundation models and adapted through sequential fine-tuning, making continual learning an essential problem setting. However, continual learning in such generative models remains poorly understood: after a task change, what aspects of the learned distribution are most easily lost, and what replay samples should be prioritized? We address these questions through the modern Hopfield energy. Recent links between modern Hopfield networks (MHNs) and diffusion models allow analyses in MHNs to be transferred to diffusion models. We introduce intrinsic forgetting as an increase in Hopfield energy after the task change. In tractable settings in an MHN, we prove that high-energy, outlier-like samples undergo a larger energy increase than cluster-like samples, implying that samples located in sharp, isolated basins are more forgettable. We further analyze memory replay and show that replay is particularly effective for high-energy samples, enabling an energy-based selection of replay samples. We validate these predictions in experiments on MHNs and two diffusion models under continual-learning settings: Stable Diffusion and a pixel-space DDPM. In these diffusion models, Hopfield energy tracks reconstruction-based forgetting, and replay experiments reveal energy-dependent mitigation of forgetting that is consistent with the MHN analysis.

Multimodal Large Language Models (MLLMs) have increasingly localized and interleaved visual evidence for deliberative reasoning. Grounding-based approaches typically focus on regions of interest (RoIs) by injecting cropped image patches or RoI-specific features into the reasoning context. However, such designs can weaken holistic scene understanding and inter-object relations, while incurring decoding costs that scale with the number and size of RoIs. Alternatively, adaptive visual feature selection often requires fine-grained supervision or complex heuristics. To address these limitations, we propose ROVER (Routing Object-centric Visual Evidence for grounded multi-image Reasoning), a lightweight, learnable plugin for efficient global visual evidence routing. Upon each object grounding prediction, ROVER injects a step-specific token triplet to synergistically: (i) aggregate the ongoing reasoning context, (ii) distill intra-image cues into a visual working space via object-centric differential attention, and (iii) route and integrate history-aware evidence across objects and images within this space for subsequent reasoning. We integrate ROVER into Qwen2.5-VL-7B and develop an interleaved SFT-to-GRPO training pipeline. Strictly adhering to the original datasets and evaluation protocols, our method achieves the best performance on MM-GCoT (+4.8% answer accuracy, +14.6% grounding accuracy) and VideoEspresso (+8.6% answer accuracy). The VideoEspresso-trained model demonstrates strong transferability, outperforming the base model by +4.7% on average across diverse benchmarks.

Speculative decoding accelerates Large Language Model inference via a draft-then-verify paradigm, yet the output projection layer becomes a bottleneck as vocabulary sizes scale. While existing static pruning methods effectively reduce this overhead, they suffer from precipitous drops in acceptance rate in specialized domains or topic-switching scenarios due to their inability to capture dynamic distribution shifts. To address this, we introduce EvoSpec, a framework that enables real-time evolution of the draft model through dynamic vocabulary and parameter adaptation. Unlike static or purely retrieval-based approaches, EvoSpec employs a context-aware mechanism that retrieves critical long-tail tokens via efficient semantic and statistical indexing. Furthermore, we propose a lightweight online alignment strategy utilizing curriculum learning to continually minimize the distributional gap between the draft and target models. Extensive evaluations across specialized domains (coding, law, and medicine) confirm that EvoSpec overcomes the limitations of static baselines. On EAGLE-3, it achieves a 1.13x speedup in these settings over the state-of-the-art static baseline FR-Spec, with 27\% lower memory overhead than standard online adaptation.

Diffusion models promise efficient parallel text generation but rely on bidirectional attention, creating a structural mismatch with pre-trained Autoregressive (AR) models. This incompatibility precludes reusing robust AR priors, necessitating prohibitive pre-training from scratch. To bridge this gap, we propose FLUID, a framework that efficiently adapts AR backbones to the diffusion paradigm. By enforcing Strictly Causal Alignment, FLUID enables seamless initialization from standard GPT-style checkpoints, circumventing the need for massive pre-training. Furthermore, we introduce Elastic Horizons, an entropy-driven mechanism that dynamically modulates denoising strides based on local information density rather than fixed schedules. Experiments demonstrate that FLUID achieves state-of-the-art performance while reducing training costs by orders of magnitude, effectively reconciling established AR foundations with efficient parallel generation. Our code is available at https://github.com/Oli-lab-nun/FLUID/tree/main.

Accurate medical coding requires consulting authoritative resources such as the ICD tabular list and coding guidelines. Existing LLM-based automated methods largely rely on LLMs' internal knowledge, which is prone to hallucination and cannot keep pace with guideline updates. We introduce RAG-Coding, an agentic, training-free method that augments LLMs with structured external knowledge: the tabular list is encoded as a knowledge graph capturing hierarchical and instructional code relationships, and the guidelines are distilled into concise, code-specific summaries rather than retrieved as raw text. To enable our study, we also introduce MDACE-2025, expert re-annotations of the MDACE dataset under the 2025 ICD-10-CM/PCS guidelines, adding code sequencing and justification comments. On MDACE, RAG-Coding outperforms the best LLM-based baseline by 3--13\% in micro-F1 across five LLM backbones, and achieves comparable micro- and macro-F1 to the supervised state-of-the-art, with higher recall ($+$11\%) at the cost of precision ($-$6\%). On MDACE-2025, RAG-Coding outperforms all baselines, demonstrating effective generalisation to updated guidelines. Ablations confirm stepwise gains, highlighting the importance of integrating structured external knowledge for LLM-based medical coding.

Humans are the bottleneck in building and improving AI. Both the models and the agents that wrap them are written, tuned, and corrected by people. The long-horizon goal of an AI that can figure out how to improve itself remains open. Two largely disjoint research lines attack this bottleneck. The harness-update school has a meta-agent rewrite the scaffold of a task-specific agent (its tools, prompts, retry logic, and search procedure) while the model weights are held fixed. The test-time training school uses hand-written RL pipelines to update the model's own weights on task feedback while the harness is held fixed. These two silos operate in isolation. We propose SIA, a self-improving loop in which a language-model agent (the Feedback-Agent) updates both the harness and the weights of a task-specific agent. We evaluate across three contrasting domains: Chinese legal charge classification, low-level GPU kernel optimisation, and single-cell RNA denoising. Combining both levers outperforms scaffold iteration alone on all three benchmarks. SIA-W+H achieves 25.1% over prior SOTA on LawBench, 12.4% faster GPU kernels than prior SOTA (1,017 vs 1,161 μs), and 20.4% over prior SOTA on denoising. Harness updates make the model agentic, shaping how it searches and acts, while weight updates build the domain intuition that no prompt or scaffold can instil.

Charts are widely used to present complex data for analysis and decision making. Existing chart understanding benchmarks mainly focus on static charts, but real-world charts are often dynamic and interactive. Key information may only appear after actions such as hovering, clicking, zooming, or dragging. Dynamic chart understanding therefore requires models to identify visible content, choose proper interactions, and reason over changing chart states. To evaluate this ability, we propose ChartAct, an interactive benchmark for dynamic chart understanding. ChartAct collects and filters 673 dynamic charts from 8 real chart websites, covers 7 common chart types, and constructs 1,440 high-quality question-answer samples. Each sample is instantiated in two environments, Dynamic Chart and Dashboard Chart, to evaluate dynamic chart understanding under different contexts. Based on ChartAct, we systematically evaluate 11 advanced multimodal models and GUI agents. Experimental results show that existing models still have clear limitations in dynamic chart understanding. The strongest model, Claude-Opus-4.7, achieves an average success rate of 84.5\%, while most models remain below 60\%. We also conduct detailed failure attribution and case analysis. ChartAct provides a new benchmark for studying chart understanding in real interactive environments. Codes at https://github.com/wulin-wulin/OSWorld_Chart

Multilingual fact verification requires evidence that is both relevant and sufficiently complete for reliable factuality prediction. However, existing systems often rely on search snippets, sentence-level evidence, or locally segmented passages, which can miss decisive context and produce fragmented evidence. To overcome these limitations, we propose SEEK, a Semantic Evidence Extraction with an adaptive chunKing framework that constructs coherent evidence chunks from full fact-checking articles by identifying semantic topic transitions and preserving local verification context. The constructed chunks are encoded using a multilingual encoder and then multilingual LLMs are finetuned using LoRA adapter for veracity prediction. Experiments on X-FACT and RU22Fact show that SEEK improves macro-f1 by up to 10% over semantic chunking, 19% over sentence chunking, and 20% over search-snippet baselines. Evidence completeness and significance analyses further show that SEEK preserves richer verification context and enables more reliable multilingual fact-checking.

Generating high-quality, pedagogically useful questions from lecture slide decks is difficult because important instructional content is distributed across both text and visual elements, and because useful questions must be scaffolded across the flow of a presentation rather than generated slide by slide in isolation. This paper describes Slide Deck Q\&A Quality Assurance (slidesqaqa), a Flask-based software system that extracts text and rendered images from PDF slides and processes them through a four-stage large language model pipeline comprising window planning, deck synthesis, slide annotation, and reconciliation. The system reasons jointly about slide modality and pedagogical role, allocates bounded question budgets, and revises draft annotations at the deck level to reduce redundancy and improve coverage. The final output is a structured JSON annotation containing deck-level goals, section structure, slide-level summaries, question sets, and evaluation scores. Initial experiments on two technical lecture decks indicate that the pipeline can filter non-instructional slides and produce high-fidelity, pedagogically coherent questions for visually complex content. The working system is at https://slidesqaqa-974767694043.us-west1.run.app The software repository is at https://github.com/blinding2submit/slidesqaqa