
ALT Visual grounding benchmarks usually do not consider the contextual uncertainty of real-world scenarios. In this study, we construct a novel benchmark, the Japanese evaluation for Ambiguous visual Grounding with Dialogue (JAG-D) dataset. It consists of human-robot conversational multimodal interactions covering different types of ambiguity, arising from visual and linguistic categories. We evaluate the capabilities of different state-of-the-art Visual Language Models (VLMs) for ambiguity recognition and resolution. We compare dialogue context grounding performance for visual and linguistic ambiguity resolution. Our experiments show that VLMs can substantially improve the resolution of linguistic ambiguity, especially Japanese zero anaphora, by using dialogue context. We further conduct a comparative analysis to understand whether grounding improvements arise from the semantic content of the dialogue or from its conversational structure.

ALT A prescription medicine in the United States comes with labeling written for prescribers, and many medicines also carry a text written for patients: a Medication Guide or a patient package insert. The two texts describe the same medicine for two readers, yet public resources rarely hold them side by side. I present PlainPairs, an open corpus built from the openFDA drug label files. Version 0.1.0 holds 35,177 prescription labels that carry a patient text, 21,416 distinct patient texts, 250,829 patient-text segments matched to prescriber sections through their regulatory headings, reading-level measures on both sides, and rule-based counts of plain-language wording. A second source, the National Cancer Institute's cancer information summaries, adds 756 summaries in health-professional and patient versions, in English and Spanish. Measured on the structure of the label, the median Medication Guide reads at grade 8.7 by Flesch-Kincaid and at grade 11.1 by SMOG. The same guides measured on

ALT After a large language model (LLM) agent generates an artifact, users often send edit requests for it. This paper studies whether an LLM can anticipate these requests before they are sent. Such anticipation is useful in practice because it reduces the effort of user review and feedback. We formalize this task as edit recommendation and build EditRecBench, a benchmark for evaluating the performance of LLMs on this task, from real-world conversation logs in WildChat-4.8M. Through benchmarking and failure analysis, we identify a practical challenge that some cases are inherently difficult to predict in real-world settings. To address this, we introduce decision model-based gating to filter out cases where predictions are unlikely to be correct before they are presented to the user. Our trained model outperforms existing decision models and raises pass@5 among shown suggestions by 12β25 points at a gating threshold. Our code is available at https://github.com/ntt-dkiku/llm-edit-recommendat

ALT Training coding agents for long-horizon tasks requires environments with verifiable feedback. Existing pipelines often follow a linear process that ends with validation, establishing usability but providing little guidance on improving training value. Recursive self-improvement (RSI) often requires model rollouts or model training followed by downstream evaluation, with costs and feedback delays amplified on long-horizon tasks. How can we use timely, reliable proxy signals to improve training value and synthesize environments at scale? We present RepoFactory, an automated pipeline that converts real repositories into verifiable environments where agents build complete projects from scratch given a language specification. Its Environment Evolution Loop optimizes a multiplicative objective combining validity, structural coupling, and implementation scope through proxy feedback. RepoFactory produces 10,159 environments across 17 software domains. Training Qwen3.5-35B-A3B on these environm

ALT Recorded sessions between users and coding agents offer a way to evaluate how agents complete real-world tasks and interact with users. However, these sessions cannot simply be replayed with new agents, because recorded user replies may no longer apply when an agent takes a different trajectory. We introduce SWE-WhatIf, which replays recorded sessions with new agents by modeling the user's intents, the criteria for satisfying them, and the conditions under which they arise. Using this model, a user simulator generates replies that adapt to unseen agent turns while remaining grounded in the recording. Across 89 real-world sessions, SWE-WhatIf generated the replies most similar to those of real users and matched the strongest baseline when sessions were rewritten. To test adaptation, we replay sessions with agents instructed to avoid or repeat actions that the recorded user pushed back on. SWE-WhatIf produced the highest win rate for avoiding these actions (66.3%), significantly above ch

ALT Recent self-improving coding agents have achieved strong benchmark performance by iteratively modifying their own executable implementations. A common pattern uses downstream task-derived evidence to decide which candidates or lineages continue evolving, which we term evaluate-and-select. Obtaining this signal requires repeatedly executing evolved agents on downstream tasks, so broader exploration incurs increasing evaluation cost, while repeated optimization against a fixed evaluation distribution may encourage benchmark-specific adaptation. We ask: can coding agents self-improve without downstream task evaluation inside the evolution loop? We introduce Explore-then-Integrate (ELITE), which integrates implementation mechanisms rather than selecting a single exploratory child to continue the lineage. At each generation, ELITE proposes diverse focused improvements, implements them independently from the same parent, and inspects the resulting programs to identify complementary, redundan

ALT Multimodal large language models (MLLMs) now assist users across sessions of images and text, carried by an external memory of what was said and shown. When such an assistant answers wrongly, the failure usually began earlier, in how a record was written, updated or retrieved, and the answer alone does not say where. We formalize multimodal memory as four stages, Write, Update, Retrieve and Answer, and define nine error types across them. A stage-resolved analysis on Mem-Gallery traces most failures to a source stage, and most of those lie before the answer. Injecting known errors against fact-preserving control edits, we then follow each tainted record downstream. Early errors survive to the answer as apparent evidence, and hurt most when no other record holds the fact. A diagnosis that trusts the quoted evidence therefore blames the stage where the error surfaces, not the one where it began. We propose MemDoctor, in which the MLLM that built the memory diagnoses its own failures. It

ALT Autonomous voice agents now conduct routine clinical follow-up calls, so some automated triage is unavoidable. A patient becoming frustrated is one thing that triage should catch. These systems typically read only the transcript the speech-to-text stage produces, not the voice, and whether that is enough is largely untested on real clinical speech. We study per-turn frustration detection on real patient calls from a deployed clinical voice agent: 150 turns from six calls labelled by three annotators, with a length-matched acted corpus as a control, and 714 turns from 30 calls. Patient turns are overwhelmingly short: across the deployed calls, roughly 59% are one or two words, against 13% in the full acted corpus, and a corpus accuracy gap remains even after the acted set is length-matched. On one-word turns, transcript-only detection does not exclude chance. Across five audio-capable language models, the transcript flags 3 or 4 of the 39 frustrated one-word turns in the 30 calls, and a

ALT Enterprise AI can now perform analytics, reporting, and data-driven workflows, yet the data foundations it relies on remain hard to operationalize: business meaning, relationships, metric definitions, and reporting conventions are absent from database schemas and scattered across undocumented systems and tribal knowledge. We introduce the Living Context Graph (LCG), a continually learned context layer that makes existing enterprise data operational for AI without requiring a copy or transformation of the underlying warehouse. LCG bootstraps base semantics and relationships from the data itself, builds a concept graph that indexes business context as queries are served, writes durable semantic learnings from verified outcomes, and withdraws or re-earns context when later outcomes contradict prior routes. We evaluate LCG through text-to-SQL, a core data-reasoning task, on a transfer-aware split of three BIRD databases comprising 499 questions: 353 form the learning history and 146 are he

ALT AI agents are being deployed on Indian workflows such as payments, tax, forms, travel and customer support in Hindi and Hinglish. Public agent evaluations rarely cover these workflows, and many score with another language model. We present Navital Gym, a set of six open environments with 321 public tasks in five task families: trip planning, UPI payments and disputes, GST invoicing and returns, form filling (including a multi-document split) and customer support in Hindi, Hinglish and three experimental Indian languages. Every task is graded by named, deterministic rules that recompute the outcome from a synthetic sandbox world. No language model judges any score. We report one run per task for three models on all six environments, Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 4.5, and for Gemini 3.1 Flash-Lite on two (UPI and the Support core). Opus solves 315 of 321 tasks, Sonnet 302 and Haiku 205. Trip planning and Hindi/Hinglish support separate the models most. We also descr

ALT Agents are increasingly built to learn from experience: they keep conversation histories, memory files, retrieved notes and skills, and reuse what worked before. We show that this experience turns against them when their tools change silently. We test a range of models with tools defined as in the Model Context Protocol (MCP). After a silent change, experienced agents fail on most tasks, while agents new to the task mostly succeed. We call this failure Experience Rot. Experienced agents copy their earlier call instead of checking the current interface, even when the tool definition shows the change. Then they report success. When the documentation is incomplete, experience is what makes agents succeed. After a change, the same experience leaves them worse off than no experience at all. The failure appears in every form of memory we test and grows with the amount of experience. Simple remedies fall short: announcing the update barely helps, and pointing the agent to the documentation fa

ALT Classifying the material facts, submissions, judicial reasoning, and the court's decision (rhetorical roles) in case law is a prerequisite for high-quality performance across virtually all NLP-in-law tasks, including legal information retrieval, summarisation, and LLM-based analysis. These rhetorical roles often intertwine and shift sentence by sentence, making reliable classification difficult. This paper introduces the first sentence-level annotated dataset of Dutch judicial decisions, labelled across four legally meaningful categories: material facts, submissions, judicial reasoning, and the court's decision, alongside non-annotated (None) content, and compares several classification strategies, including a staged pipeline that separates None detection from the rhetorical roles, and a Bayesian update that leverages information about the section-header a sentence belongs to. Our results show that no single strategy dominates on both precision and recall: header priors help most when

ALT Disentangling writing style from semantic content is a fundamental challenge in literary text modeling. Style-content entanglement causes models to rely on semantic content for authorship attribution (AA) tasks and memorize author-specific content for style imitation (SI) tasks. We propose EDITORS (EditingLoRA Subspaces), a diagnose-then-deflate framework that adapts pretrained large language models and prioritizes stylistic features, making it effective for both AA and SI tasks. EDITORS first trains a diagnostic LoRA adapter using style-neutral statements derived from the training corpus to identify a content-oriented subspace. It then trains the final style adapter under activation-space deflation that projects input activations away from the identified content-oriented directions. We empirically demonstrate that EDITORS is effective. On three AA benchmarks spanning literature, social media, and news, EDITORS achieves state-of-the-art accuracy, outperforming the strongest existing ba

ALT A system that reads nothing and always answers the most frequent letter scores BaselineP on our Hausa medical multiple-choice set. NBelowBaseline of the NModels small open models we evaluate do not beat it. We lead with that because the usual summary of a multilingual result β a percentage-point drop from English β does not say whether what survives the drop is performance at all. On a paired EnglishβHausa set of NEval MedMCQA items, three instruction-tuned models of 3β4B parameters lose between GapMedGemma and GapQwen points of accuracy, every gap significant after Holm correction. Back-translating the Hausa items into English recovers between half and two-thirds of the gap for the two general-purpose models but not demonstrably for the medically adapted one, so the deficit is substantially, though not wholly, a matter of the language of presentation. Answers are also far less stable: agreement across three option orderings falls from ConsEnQwen to ConsHaQwen in the worst case. An exp

ALT Text-to-image (T2I) safety evaluation typically assigns each prompt or image a single culture-blind harm label, so content that is harmful only within a particular culture can pass as benign. We introduce CultureCraft, an agentic framework that constructs T2I prompts that appear benign under a general-purpose rubric yet carry a latent harmful meaning in a target culture, and use it to build CuRSE-Bench: 10,709 culture-specific prompts and 2,648 culture-general controls over six regions and ten harm axes, each drawn from a predefined knowledge base of documented harm vectors and traceable to at least two independent sources. Scoring every prompt under six region persona judges and a paired culture-blind judge, we find that the culture-blind judge under-rates culture-specific latent harm while matching the persona judges on the controls. Each persona judge, in turn, chiefly decodes its own region's latent harm under our instrument: a selective cultural blind spot that persists in deploye

ALT Digital assessment traces are increasingly used to predict student performance, often evaluated by correlating predicted and observed scores pooled across tasks. However, such correlations may reflect between-task differences rather than performance differences among students within tasks. To investigate this, we analyzed up to 527 browsing sessions from 424 university students completing open-Internet Critical Online Reasoning tasks and assessed the cross-task stability of their ratings. We evaluated the same out-of-fold predictions within and across tasks, compared three validation tiers up to leave-one-task-out, and tested incremental predictive contributions of consultation order and relational structure against matched permutation and randomization controls. None of the twelve prespecified primary tests was significant: on held-out tasks, consulted-page features showed no significant association with relative student performance, nor did observed order or relational structure add

ALT We introduce Nemotron 3.5 Bolt E4B 4GB, an open and efficient small language model further post-trained from Gemma-4-E4B-it and optimized for speed and intelligence per GB of memory used for fast conversational agents. The model carries forward Gemma's strong conversational capabilities and further improves multi-turn tool use, instruction following, and input grounding for fast, local voice/device assistants, interactive personas, and embedded/Edge interfaces. The model combines natural conversation with multi-turn agentic capabilities to power voice/device assistants, interactive characters and robotic/IoT interfaces running on GeForce RTX, Jetson, and RTX/DGX Spark. Experiments on 21 benchmarks show that Nemotron 3.5 Bolt E4B 4GB achieves the highest average accuracy while generating the fewest output tokens in both reasoning and non-reasoning modes, compared with Gemma-4-E4B-it and Qwen3.5-4B. The model is released with a DFlash drafter. With DFlash, Bolt E4B achieves 2.74Γ the per

ALT Agentic memory systems give LLMs persistent memory by extracting experience into text and retrieving it at query time, but the model itself does not evolve: however many interactions it has had, it starts each query the same. Extraction errors compound, implicit cross-session patterns that no single fragment states cannot be surfaced, and the construction cost grows with every conversation. Fine-tuning on past interactions does change the weights, but requires backpropagation at serving time and degrades on long histories. We introduce CLAM (Carried Latent Adapter Memory), which combines the strengths of both: a recurrent hypernetwork performs parametric consolidation, compiling a stream of session activations into a low-rank adapter in a single forward pass with no gradient steps and no language-model generation calls, while a complementary episodic buffer supplies verbatim detail at inference. A gated latent state carries knowledge across sessions in fixed size, so the parametric mem

ALT Process Reward Models (PRMs) are widely used for step-level evaluation of multi-step reasoning in large language models, but existing methods are often limited by weak historical dependency modeling and insufficient global awareness. They typically fail to leverage prior step evaluations for coherent judgment and lack visibility into subsequent reasoning, leading to suboptimal assessment of intermediate steps. To address this issue, we propose the Global Perspective Process Reward Model (GPRM), which jointly models the problem, previous reasoning steps and their evaluations, the current step, and future steps to enable history-aware and globally informed process evaluation. We further introduce a structured evaluation framework along four dimensionsβlook back, look ahead, self-check, and goal alignmentβto assess prior consistency, future plausibility, step validity, and task alignment. Based on this framework, we construct datasets for supervised fine-tuning (SFT) and reinforcement lea

ALT Evaluating product advice from a language model requires assessing whether the recommendation meets the user's needs and identifying where the model falls short when it does not. However, matching a product mentioned in an online reply cannot answer either question on its own: several products may be suitable, and the recorded answer may itself be poor advice. When shown a community-listed alternative alongside their own sampled car recommendation, six models prefer the alternative in 1.9 to 6.5 percent of eligible disagreements, consistently across both source-blind presentation orders. We propose a diagnostic framework that separates the user's stated needs, the model's added assumptions, candidate generation, filtering, and final choice, with 19 possible failure modes and the evidence needed to assess them. We apply this framework to product-advice data and to a separate product-selection task with externally assigned relevance labels. The experiments distinguish naming a community-
