Papers

Positional embeddings (PE) in Transformers encode token distance and order but are largely agnostic to syntactic structure. We introduce Syntax-informed Positional Embeddings (SiPE), which learns a lightweight syntactic prior from dependency parses during pretraining and injects it across all three dominant PE families (absolute, relative, rotary), for both encoders and decoders, leaving self-attention and the rest of the architecture untouched. We isolate where and how the prior should enter the model, and find it depends on the architecture: for autoregressive decoders that use relative PE, the prior is strongest when coupled multiplicatively with the relative-position term of the attention score, outperforming injection into the input embeddings, into self-attention, or into the positional and attention terms jointly—while for encoders it is best added directly to the input embeddings, composing with each encoder's native positional mechanism. We find that models pre-trained with SiPE improve on the SyntaxGym benchmark by up to $10.3\%$ while simultaneously reducing perplexity by $9.0\%$ over a base model with no syntactic supervision—a metric nearly every existing syntax-injection method instead degrades. Crucially, these gains extend beyond syntactic generalization: SiPE also improves real-world language understanding, raising scores on the GLUE benchmark by up to $8.2\%$ over a model trained without it. Unlike existing syntactic language models that marginalize over many parses at inference or discard syntax at runtime, SiPE conditions on a single parse, establishing a new Pareto frontier between syntactic supervision and inference cost.

This paper presents ECHO (Enhanced Care & Health Observer), a locally-deployable conversational health assistant for long-term chronic care management. ECHO integrates three complementary software modules developed under shared supervision as a unified system. The core module is an agentic chatbot built on a ReAct loop orchestrated via LangGraph, equipped with 17 clinical tools and a temporal knowledge graph for persistent cross-session memory; it achieves a 94.9% tool-execution pass rate across a 59-scenario benchmark with GPT-5 Mini. A two-stage hybrid safety layer intercepts all incoming queries: a rule-based layer handles explicit crisis signals and jailbreak attempts in under 1ms, while a signed graph neural network (GNN) with APPNP-style propagation classifies boundary cases by clinical intent, achieving 88.8% accuracy and 90.6% unsafe recall on a 2,537-query annotated Turkish health dataset while outperforming zero-shot LLM baselines including Llama 3.3 70B. A multimodal speech assessment module combining Whisper acoustic encoding and BERT text encoding with cross-attention fusion estimates emotion, depression, and pain, reaching a mean macro F1 of 0.652. The full system is implemented as a web application that can run entirely on consumer hardware, with no patient data transmitted to external services, supporting compliance with GDPR and KVKK.

While Large Language Models (LLMs) show great promise for personalization, they often lack specialized domain knowledge. Conventional solutions like fine-tuning struggle with high computational costs and rapid data updates, while Retrieval-Augmented Generation fails to provide fine-grained, token-level steering. Furthermore, chat-based interfaces remain dominant, whereas productive co-writing paradigms have not yet been well exploited beyond the coding domain. To this end, we introduce SteerWrite, a training-free framework designed for personalized co-writing. Our method effectively adapts the base model to specialized domains without gradient updates, with specific designs tailored to small datasets. Experiments demonstrate that SteerWrite achieves state-of-the-art performance across diverse datasets, metrics, and models, significantly reducing human editing effort.

Large language models (LLMs) have been shown to exhibit strong Python preferences when generating project-level code, but there is currently no systematic way to measure this behaviour across new models. To bridge this gap, we introduce LangChoiceBench, a project-level code-generation benchmark for measuring Python preference, recommendation-implementation consistency, and language diversity. LangChoiceBench covers 28 projects across seven software areas where Python is often a poor default. We evaluate 25 diverse LLMs and find that Python remains heavily over-selected, recommendation-implementation consistency is low, and smaller open-weight models generally show stronger Python preference and lower language diversity. We further analyse 9,826 reasoning traces and find that most Python choices are automatic or driven primarily by ease, rather than explicit consideration of project requirements. In a smaller but important set of cases, models fabricate contextual support for choosing Python - a failure mode we call phantom evidence - or produce code that contradicts the language selected in their own reasoning.

In India, almost every social benefit starts with a form, yet the people who need these benefits most are often unable to read or write. Reaching them requires a spoken conversation. Today that work falls to frontline health workers who enroll beneficiaries one at a time, a poor use of stretched capacity. We built FormBharo ("fill the form" in Hindi), a voice agent that fills a structured form over a phone call under tight latency and cost budgets by pairing Large Language Models (LLMs) with deterministic, rule-based validation and flow control. It is being piloted with ARMMAN, an NGO running large-scale maternal and child mobile-health programs in India, to enroll low-income, Hindi-speaking mothers in antenatal and postnatal care. To our knowledge, it is the first voice agent piloted to fill an enrollment form for this population. We openly release FormVoiceAgentBench, a benchmark pairing human-recorded Hindi audio with 3,760 multi-turn conversation tests across 960 simulated calls, to evaluate our agent's components (transcription, extraction, reply generation) and end-to-end form completion under real acoustic variations. Form completion drops by up to ~41 points when LLMs receive error-prone real-speech transcripts instead of reference ones. The rule-based controls recover many turn-level extraction errors, helping smaller, cheaper models match or surpass frontier models on form completion. Component performance does not predict end-to-end performance: GPT-5.5 leads turn-level extraction accuracy on reference transcripts (99.8%) but ranks lower on form completion. Since errors both propagate and cancel across the pipeline, the optimal model choice of models emerges only through end-to-end evaluation. Finally, no single model is best across accuracy, cost, and latency at once, so we use a Pareto-based weighted-sum scalarization to select a deployable configuration balancing the three.

Epitopes determine where antibodies bind antigens and shape downstream therapeutic properties such as functional blockade and escape resistance, making epitope understanding central to antibody drug discovery. Although large language models (LLMs) have shown strong biomedical reasoning ability, it remains unclear whether they can infer epitope information directly from antigen and antibody sequences. Existing epitope resources typically focus on isolated prediction tasks or rely on specialized structural settings, while general protein benchmarks do not evaluate epitope-centered decisions across the antibody development workflow. To address this gap, we introduce EpiBench, a closed-book, sequence-based, and automatically scorable benchmark for evaluating epitope reasoning in LLMs. EpiBench contains 1,609 curated samples grounded in structural antibody–antigen contacts, curated functional B-cell assays, and deep mutational scanning escape measurements. It covers five connected tasks: targetable region discovery, antibody-conditioned epitope identification, epitope binning, functional epitope assessment, and antibody escape assessment, with controlled sampling to reduce shortcut-based evaluation artifacts. We evaluate nine general-purpose LLMs and analyze their behavior through task-specific baselines, antigen length stratification, explicit-reasoning comparison, and failure-mode inspection. The results show that current LLMs capture partial epitope-related signals but remain limited in antibody-specific sequence grounding, long-context residue localization, and biologically grounded reasoning. Therefore, EpiBench provides a diagnostic testbed for measuring and improving sequence-aware biomedical LLMs toward reliable LLM-assisted antibody discovery.

Much clinical value is conveyed not through structured records but through communication: exchanges in which patients describe symptoms, clinicians reason and give instructions, ambulances hand over to emergency departments, and nurses pass on a shift. Such language differs from tabular data because meaning depends on speaker role, intent, causality, uncertainty, omission, and channel noise. Healthcare natural language processing must therefore interpret information as conveyed rather than coded. This requires well-annotated corpora, which are scarce because authentic exchanges are private, fragmented, and costly to annotate. Large language models offer a way forward by transforming clinical sources, such as records, diagnostic labels, symptom lists, or care plans, into written and transcribed communication for downstream models. We present a structured narrative survey organized by source representation, communication form and participants, generation method, and downstream task, complemented by thirteen novel case studies. These build clinical NLP systems for communication channels and languages without labeled real-world data, including EMS pre-arrival reports, field-radio casualty documentation, nurse handoffs, patient-portal triage, and low-resource discharge communication. They show that synthetic communication can bootstrap such systems. Findings include the competitiveness of fine-tuned encoder models over evaluated zero-shot baselines and the value of deliberately degraded communication for robustness. The main limitation is that most studies evaluate on held-out synthetic communication, while train-on-synthetic, test-on-authentic evidence remains limited. We conclude that syn-thetic clinical communication is becoming a practical research resource; establishing it as reusable clinical infrastructure will require authentic-data transfer, safety and external validation.

LLM agents that repair failures often discard successful corrections, forcing later episodes to rediscover similar solutions. We study whether finalized repair outcomes can improve subsequent Text-to-SQL episodes without parameter updates. We introduce MERIT, a training-free agent that maintains an online dual-polarity memory of oracle-verified corrections and observed unsuccessful directions. Under oracle-assisted benchmark feedback, only memories from earlier finalized episodes are eligible for retrieval. A deterministic classifier assigns a coarse failure type, which conditions a hybrid lexical-dense retriever before the frozen model generates each revision. Using Qwen2.5-7B-Instruct with identical initial predictions and repair budgets, \method{} improves execution accuracy over stateless iterative repair from \(66.34\%\) to \(69.79\%\) on Spider and from \(47.35\%\) to \(48.44\%\) on BIRD. Paired analyses provide clear evidence for the Spider gain but weaker evidence on BIRD. MERIT is not reliably separated from untyped dynamic retrieval on either benchmark, while Reflexion-style memory reaches \(51.24\%\) on BIRD at substantially higher inference cost. Ablations show that negative memory contributes modestly, the value of type conditioning and lexical-dense ranking is dataset dependent, and schema-local experience provides the most consistent benefit. These results clarify when causal cross-query memory improves repair and when broader memory representations remain preferable. Our implementation is available here:

Mobile GUI agents can operate apps through pixel perception and touch actions, making them a promising interface for collecting and improving long-horizon mobile interaction policies. However, real trajectories are difficult to obtain for sensitive apps and privacy-critical operations. At the same time, existing simulated environments are costly to scale up, and GUI world models still suffer from unstable generation, limited modality coverage, and inconsistent action-transition logic. To address these limitations, we propose AppDeltaWorld, a transition-grounded delta code world model that predicts the next GUI as a reachable code update rather than as an unconstrained image or text description. AppDeltaWorld retrieves app-specific Level-1 HTML references under an action-transition constraint, generates Level-2 executable HTML conditioned on the current screen, action, predicted next-screen text, and retrieved structure, and inserts generated visual assets into image slots before browser rendering. As a world model, AppDeltaWorld achieves the highest fidelity on CMGUIBench-500 under Code2World evaluation, with clear gains in structural layout and UI element reconstruction over image-only and code-only baselines. As a training environment, AppDeltaWorld supports filtered closed-loop SFT data construction that, when combined with public supervision, enables AppDeltaAgent to achieve state-of-the-art performance on AndroidLens and consistent gains on MobileGym and MobileWorld. Moreover, world-model-based test-time reinforcement learning enables policy adaptation and shows further improvements without additional interaction with real apps.

Large language models (LLMs) can leave small stylistic traces in text written with their help. The most discussed is the em-dash (U+2014), especially the unspaced form word—word, which is normal in typeset English prose but unusual in U.S. press writing, where AP style calls for spaced dashes. This study asks whether that trace is measurable in congressional press releases. In a preregistered design (OSF: 10.17605/OSF.IO/U5NEY), 146,239 scraper-sourced releases from 480 House and Senate offices (2021-2025, the open congress-press dataset) were analyzed: density of unspaced prose-form em-dashes per 1,000 characters of cleaned text, Poisson/negative-binomial models with a length offset, clustering by office. Density stayed within 0.10-0.12 per 1,000 characters through 2021-2024, then rose to 0.217 in 2025, more than twice the four-year baseline; the share of releases with such an em-dash rose from ~13% to 24.8%. The primary frequency ratio (2023-2025 vs 2021-2022) was 1.55 (95% CI 1.28-1.93; exact registered cut-off: 1.528), just above the prespecified 1.5x threshold. The rise was net-new (hyphen density stable), held within authors (75.6% of 262 continuous offices increased; p ~ 1e-16) and in a closed panel of 224 offices, and survived falsification tests: three placebo cut-offs were null, the pipeline showed no step at the 2024/2025 boundary, and continuing offices carried the rise. A segmented regression finds no step at the ChatGPT cut-off but a clear post-period acceleration; the 2025 rise is symmetric across parties and chambers. Because the registered validation gate was formally breached, the full preregistered decision rule was not met; the interpretation (broad diffusion of LLM-assisted writing as the models matured) is offered as exploratory. The em-dash remains a population-level marker, not a per-release authorship detector, and the design supports no causal claim.

Existing guidance identifies excessive agency, excessive permission, weak task-bound authorization, and inadequate agent controls as important risks. Control frameworks also describe capabilities for constraining, authorizing, observing, validating, and responding to agent activity. Yet security programs still need a way to manage persistent deployed instances that span components and outlive any one event. We propose the agentic posture vulnerability (APV) as a task-conditioned vulnerability-management abstraction: a durable record for a composed agent-control exposure. One posture may produce different runtime manifestations across tasks; APV links those manifestations to the invariant posture and remains open until authority is narrowed, a missing control is added, risk is accepted, or closure is verified. APV is not proposed as a new root-cause class of risk; it operationalizes existing excessive-agency, authorization, and control-composition weaknesses. We distinguish APVs from CVE-addressable product defects, OWASP Excessive Agency, Agent Baseline control outcomes, and the runtime authorization-execution gap. We then provide a field vignette, a thresholded definition, six recurring APV patterns, a vulnerability lifecycle, a minimum record, a control-and-closure matrix, tooling implications, and a testable research agenda.

User requests serve as research specifications for deep research agents, shaping what evidence to seek and how to synthesize it. In personalized deep research, these specifications must additionally reflect user goals, constraints, preferences, and evaluation criteria. User context can be incorporated either within the deep research pipeline or into the research specification provided as its input. We focus on the latter, refining the user request into a personalized research specification before passing it to an unchanged deep research agent. This requires resolving three coupled decisions: which framing factors are relevant, whether the available user context sufficiently supports them, and whether to retrieve user memory, ask the user, or stop and refine the query. For training, G-STEER organizes framing factors as elicitation targets in an Intent Elicitation Graph that captures their dependencies. It learns a clarification policy from graph-scaffolded trajectories spanning diverse factor dependencies and evidence conditions. The policy produces a refined query while balancing target coverage against the costs of evidence acquisition. Experiments show that G-STEER achieves the strongest overall weighted target coverage and the highest downstream report personalization across both evaluated DRAs, while asking roughly one third as many user questions as a strong clarification baseline.

Standard Large Language Models (LLMs) execute layers sequentially. Dynamic layer routing, i.e. search for a different execution path through layers involving layer repetitions, skips and other moves, can improve performance. Existing routing approaches often require updating model weights, running expensive search loops per test instance, or demand ground-truth labels during inference. In this work, we propose Markov Chain Routing of Transformer Layers (MACRO), a framework that learns task-specific routes over LLM architectures without modifying underlying parameters. MACRO models layer routing as a context-dependent Markov policy conditioned on layer indices, computation budget phases, directional displacements, and operator context, supporting skip, repeat, and residual hidden-state addition operations. The Markov route distribution is updated via feedback on training data and decoded using a top-k Viterbi algorithm to isolate high-probability candidate programs. We evaluate MACRO across diverse reasoning and knowledge benchmarks on multiple open-weight LLMs. MACRO achieves a +5.0% average accuracy improvement over the unrouted baselines, with largest gains on small models. We outperform the best dynamic routing approach Dr. LLM by +7.2%, while reducing route-search time 9.4x (from 14.8 to 1.6 hours). Our code is publicly available at https://github.com/Batorskq/MACRO.

Retrieval-Augmented Generation systems rely on similarity scores to retrieve relevant content, yet scores are not directly comparable across embedding models due to differing geometric properties, complicating model migration and limiting threshold reuse. We study how similarity scores can be related by learning mappings between score distributions rather than embeddings. We introduce Synthetic Query Probing, generating queries from documents to create controlled query-chunk pairs, enabling large-scale, reference-free analysis of cross-model similarity behavior. We evaluate the approach on multiple embedding configurations and learn score conversion functions using linear, isotonic, and quantile mappings. Experiments on SciFact and a proprietary corpus show that while models largely agree on rankings, their absolute scores exhibit systematic distortions. Learned mappings partially align these spaces and improve threshold portability, with isotonic regression performing best. Our results highlight the need for cross-model calibration and position Synthetic Query Probing as a scalable framework for analyzing embedding comparability.

We present MameLoshnLM, the first open-source 8B-parameter language model built specifically for Yiddish. Despite Yiddish's rich textual tradition, its limited digital presence and the scarcity of reliable evaluation resources have constrained progress in Yiddish language modeling. Existing multilingual corpora and benchmarks are often poor proxies for the language, containing substantial amounts of noisy, machine-translated, and misclassified text. We address these gaps by introducing Oytser, a high-quality Yiddish pretraining corpus that combines contemporary web-native sources with literary materials, and Kashes, a multi-task benchmark spanning translation, linguistic analysis, information extraction, and language understanding. Using these resources, we continue pretraining Llama 3.1 8B to obtain MameLoshnLM. Across the tasks in the benchmark, MameLoshnLM outperforms open baselines of similar scale. Our analyses show that these gains are not only quantitative: relative to general-purpose multilingual models, MameLoshnLM better captures language-defining lexical and morphological patterns, pointing to a broader failure mode of noisy web-scale multilingual data for low-resource languages. Our results provide both a foundation for Yiddish NLP and a practical template for language model development in historically rich but digitally underrepresented languages.

Large language models (LLMs) excel in structured tasks but struggle with dynamic social interactions, where success requires long-term goal coordination and rapid adaptation. Current methods often apply uniform goal-based rewards to every utterance, overlooking the specificity of objectives at each dialogue turn and failing to account for the rationale of potential strategies. Inspired by the Theory of Planned Behavior, we propose the Think-Strategy-Response (TSR) framework, which decomposes social dialogue into two hierarchical stages: high-level strategic planning and low-level linguistic execution. To optimize TSR, we introduce Linearized Hierarchical Reinforcement Learning with Variance-Gated Rewards (LHRL-VGR), a novel algorithm that dynamically routes rewards - balancing goal completion and strategy adherence - based on the variance of goal achievement scores. Experiments on the SOTOPIA benchmark show that our approach fine-tunes a Qwen2.5-7B agent to surpass the GPT-4o baseline by 7.32% in goal completion success, demonstrating state-of-the-art performance in multi-agent social negotiation tasks.

Human social communication, such as affection and intent, is often conveyed in highly implicit ways, where underlying meanings are expressed through indirect, socially and culturally grounded signals rather than explicit statements. Such implicit social contexts are pervasive in real-world interactions, yet there remains a lack of a formal and systematic framework for studying them. In this paper, we introduce Implicit Social Context Analysis (MoCA), a novel task that systematically models implicit social scenarios along three key dimensions: affection, intent, and stance. We construct a high-quality benchmark containing 3,108 multimodal instances collected from real-world sources, with fine-grained cognitive annotations revealing who expresses what toward whom, as well as how and why it is conveyed. Using the MoCA dataset, we show that state-of-the-art multimodal large language models struggle significantly with this task because of their reliance on explicit cues and limited ability to reason over latent social contexts. To address this challenge, we propose Conflict-Driven Abductive Reasoning (CoDAR), a novel framework that models the discrepancy between observed expressions and expected truthful behavior as cognitive conflict, thereby enabling the inference of hidden mental states. Extensive experiments demonstrate that CoDAR substantially improves model performance. Nevertheless, a large gap from human reasoning remains, highlighting the fundamental difficulty of implicit social understanding.

The reliability of Large Language Models (LLMs) is often compromised by factual inconsistencies, including hallucinations—cases where generated content is not supported by the underlying source. We present HallDetect, a lightweight, reference-free, and black-box framework for hallucination detection that we evaluate not only on summarization but across a broader range of source-grounded generation settings. HallDetect builds on decomposition-based factuality evaluation: generated content is decomposed into atomic claims, each verified by a compact encoder-based entailment model through a contrastive formulation over a multi-scale library of source chunks, and aggregated with an asymmetric score in which a single confidently contradicted claim flags the response. Under a controlled protocol in which all methods share the same 4-bit quantized backbones and consumer-grade hardware budget, HallDetect outperforms comparably resourced generative and embedding-based baselines on three of four benchmarks while remaining stable across backbone families, and yields a claim-to-span audit trail that localizes each error.

Multimodal metaphor understanding requires identifying cross-domain Target-Source mappings and inferring sentiment from visual and textual evidence. However, existing benchmarks largely evaluate these subtasks separately and lack evidence-grounded explanations, limiting their ability to assess coherent cross-modal reasoning. To enable unified and evidence-grounded evaluation, we introduce M$^3$R-Bench, a benchmark of 1,000 human-annotated image-text instances with joint labels for metaphor occurrence, Target-Source mapping, and sentiment, alongside stage-wise explanations for evidence identification, mapping establishment, and sentiment inference. Evaluations reveal that existing models frequently rely on textual shortcuts and produce Target-Source mappings inconsistent with visual evidence. To address this evidence-mapping mismatch, we propose M$^3$R-Reasoner, which combines curriculum-based reasoning supervision with task-aware reinforcement learning. Built on an 8B-parameter backbone, M$^3$R-Reasoner achieves 37.90% Joint Label Accuracy (JLA) metric, outperforming the strongest proprietary baseline by 19.50 percentage points. It further surpasses GPT-5.5 in image-text evidence and sentiment justification scores by 28.45 and 28.11 points, respectively. The dataset and code are available at https://github.com/hongshi4/m3r-bench-artifact.

Self-evolving agents accumulate capability by distilling reusable skills from their execution trajectories, but we find this process is not monotonic: past a critical pool size, newly added skills degrade performance instead of improving it. We formalize this capability-contamination phase transition and trace it to a structural cause: once a defective skill enters the decision context, it becomes reference material for distilling later skills, forming cross-round contamination chains. We further show the contamination is structurally irreversible: removing a source skill after the fact cannot erase the flawed reasoning its descendants have already inherited, so post-hoc rollback recovers only a small fraction of the lost performance. This makes skill admission a pre-commit necessity rather than a post-hoc fix, and motivates Verifier-as-Gatekeeper (VaG): a progressive trust hierarchy whose three heterogeneous critics - structural validity, behavioral harmlessness, and semantic consistency - filter each skill individually, coupled with a marginal-gain subset selection that removes combinatorial contamination at the top tier before skills reach the runtime context. On Terminal-Bench 2, unconditional accumulation rises to a peak and then degrades, giving back most of its gains as the pool keeps growing, and post-hoc removal of the culprit skills recovers only a small part of the drop - the empirical signature of irreversibility. In contrast, VaG improves every round, reaching 72% pass@1 with a pool roughly 5x smaller, and its frozen skill pool transfers positively to four other backbones and a second benchmark without re-evolution. Ablations confirm the three critics are complementary and mutually non-substitutable, each intercepting a largely disjoint class of harmful skills.

While standard Next-Token Prediction (NTP) lays the foundation of language model pre- training, its teacher-forced training paradigm may not be optimal for long-horizon reasoning and planning. Recent works such as Multi-Token Prediction (MTP) and Next-Latent prediction (NextLat) try to mitigate the problem through predicting multiple future tokens and self-supervised prediction in the latent space. However, those auxiliary objectives either have a limited horizon or suffer from compounding error from multi-step rollout. We introduce Hierarchical Latent Prediction (HiLP), which introduces an auxiliary higher-level abstract latent to help reduce the error accumulation effect in latent-space rollouts. Experiments show that HiLP can lead to longer-horizon coherent belief state representation and demonstrate the effectiveness of our method across coding and multi-step reasoning benchmarks, and offers more speculative decoding efficiency.

On-Policy Distillation (OPD) is emerging as a promising alternative to reinforcement learning for LLM post-training, yet its effectiveness in multilingual settings remains underexplored. We study OPD and its advanced variant, On-Policy Delta Distillation (OPD$^2$), for mathematical reasoning in English, Korean, and Japanese. OPD$^2$ improves OPD by using the probability gap between a post-trained teacher and its base model as the learning signal. Experiments with Qwen3 show that OPD$^2$ consistently outperforms the original OPD, with particularly strong improvements in Korean and Japanese, and generally narrows the English-Korean performance gap. We further find that English-only OPD can also increase performance for Korean and Japanese, but often shifts the responses toward English, highlighting the importance of multilingual data to preserving target-language responses.

A fundamental challenge in evaluating and training autonomous agents is measuring the intrinsic difficulty of the tasks they attempt. Estimating this quantity can be useful for environment designers creating synthetic data or benchmarks, as well as for constructing training curricula, thereby offering a way to reduce compute costs. Such an estimation becomes increasingly important as agents (particularly LLM-based) move into longer-horizon domains, where empirical trial-and-error becomes a severe computational bottleneck. However existing work is largely confined to static tasks and relies on evaluation metrics that, as we show, can give an incomplete picture of predictive performance. In this paper we study ex ante difficulty prediction across 17 agentic benchmarks spanning coding, mathematics, machine learning, web navigation, function calling, and other domains. We find that AUC can mask poor difficulty estimates, identify token-level entropy as a useful predictive signal, and demonstrate how residuals between expected and observed difficulty can expose environment flaws such as contamination and infeasibility.

Multilingual text embedding models are commonly adapted using a single training objective across diverse tasks, despite different tasks requiring fundamentally different optimization strategies. We introduce Task-Conditional Flow Matching (TCFM), a multilingual embedding adaptation framework that selectively applies Flow Matching to translation tasks while optimizing retrieval, classification, and pair-classification tasks with objectives better aligned to their learning dynamics. TCFM further combines teacher-guided representation preservation with a three-stage curriculum to enable stable adaptation. Evaluated on the Indic Massive Text Embedding Benchmark, TCFM establishes a new state-of-the-art, consistently improving embedding quality across a diverse set of multilingual tasks and generalizing across embedding model families. We will publicly release the codebase and datasets upon acceptance of the paper.

Machine unlearning has become a critical capability for safely removing specific, sensitive knowledge from large language models (LLMs). Current state-of-the-art approaches primarily rely on iterative, training-time unlearning via fine-tuning. However, even when utilizing parameter-efficient dimensionality reduction techniques like LoRA, gradient-based optimization remains computationally expensive and lacks explicit analytical formulations. It can also leave the targeted knowledge merely hidden rather than removed, to the point that simply quantizing the unlearned model restores much of what it was supposed to have erased. To resolve this, we propose a novel one-shot unlearning approach, abandoning iterative optimization in favor of a direct, exact analytical solution. We frame the unlearning process as a ridge-regularized least-squares optimization problem, deriving a closed-form additive update for targeted weight matrices. This update forces the selected layer to suppress unwanted content while strictly preserving its behavior on retained data. Computed from gradient-free forward passes alone, with no backpropagation and no iteration to convergence, GROM applies the weight edit in mere seconds, which makes it orders of magnitude faster than traditional fine-tuning. Extensive evaluations demonstrate that GROM achieves state-of-the-art forgetting-utility trade-offs on TOFU-5%, TOFU-10%, MUSE-Books, MUSE-News and WMDP, significantly reducing computational overhead without sacrificing overall model performance. Because the update removes the targeted content from the weights instead of masking it, GROM also withstands the low-bit quantization attack that recovers much of the content a gradient-based baseline had appeared to forget. Our code is publicly available at https://github.com/Batorskq/GROM.

Recognizing new and rare words - named entities, acronyms, domain specific special words, and other items scarce in training data - remains a key challenge for automatic speech recognition (ASR). We compare two strategies for this: context biasing methods, where an ASR model is extended such that during inference a word list can be supplied, and speech large language models (LLMs) prompted with context directly. We evaluate two context biasing methods based on Whisper against three speech LLMs across read and non-read speech, reporting biased, unbiased, and overall word error rate (WER). The context biasing methods cut biased WER by up to 88% relative while leaving other words largely unaffected. Speech LLMs excel on read speech but generalize less well to non-read speech, and prove sensitive to distractor count and prompt word order. We characterize the resulting trade-offs to guide method selection.

Large language models (LLMs) can generate fluent and convincing text at scale, creating growing risks for misinformation dissemination, educational misuse, and platform governance. These concerns make robust detection of machine-generated text increasingly necessary. Recent zero-shot detectors mainly exploit probability-based statistical discrepancies, but they do not explicitly account for the training process of LLMs, which leaves a distinct generation mechanism insufficiently modeled and limits detection robustness. To address this issue, we propose EchoPrompt, a training-free detector based on latent prompt restoration. Our key intuition is that machine-generated text is typically produced conditioned on an upstream prompt, and this hidden dependency can be partially reactivated by prepending a unified generic prefix. Specifically, EchoPrompt restores a generic assistant-response context, measures the induced likelihood gain with an instruction-tuned model, calibrates it against the corresponding base model, and aggregates the resulting differences into a score that quantifies latent prompt dependency. Extensive experiments show that EchoPrompt achieves state-of-the-art performance among zero-shot detectors while maintaining strong robustness across challenging evaluation settings.

As capabilities rapidly increase, AI agents can move from running inside one app to acting across a user's devices over time. Yet existing agent systems still fall short in this scenario. This is because observations are scattered across devices and moments, but mainstream systems are not designed around this fact: a single agent that treats devices as tools lacks effective state management for all devices across time, and multi-agent systems coordinate across agents but do not maintain the compact carried state a cross-device, cross-time request needs. We argue that the agent should maintain an effectively designed state that organizes engagement evidence, stated facts, and the standing request in a compact, action-ready form for deciding its action given the current observation. To compare state designs, we construct a benchmark of user-agent interaction across devices and time. We instantiate this principle in Unified Agent, a stateful agent that carries interaction evidence across devices and moments and uses it with the current observation to act. In the default setting, it significantly outperforms our adaptations of four published designs. Across changes in multimodal large language model (MLLM) family, capability, and reasoning effort, it remains ahead of all compared systems, demonstrating that the state-design advantage is robust across MLLM settings. Our code and data will be publicly available on GitHub.

Large Language Models (LLMs) are often used as evaluators of text quality, known as LLM-as-a-Judge, which can outperform conventional automatic evaluation metrics that rely on reference texts. However, LLM evaluators tend to generate particular scores regardless of the context of the evaluated text, which is known as scoring bias. This study proposes a novel method to mitigate this scoring bias. An LLM is instructed to randomly generate number tokens, and the latent numerical bias of the LLM is identified by measuring the deviation of the observed distribution of numbers from the uniform distribution. A definition of a downstream task, for which an LLM evaluator is used, is added to the prompts for random number generation to measure task-specific latent number bias. In the evaluation by an LLM, the token generation probabilities for a given input are rectified considering the LLM's latent number bias. Results of the experiment on four different tasks, evaluation of LLM alignment, evaluation of summarization, Semantic Textual Similarity, and Semantic Textual Relatedness, demonstrate that our proposed method outperforms the baselines, including an LLM without debiasing and previous calibration methods. In addition, it is confirmed that scoring bias varies across LLMs, tasks, and score ranges, indicating the importance of measuring latent number bias as the case may be.

We study a specific sparse post-processing pipeline for Random Indexing (RI) on kinship analogies in a small fairytales corpus. The published artifacts use uniform RI context accumulation with 200 dimensions and eight nonzeros, followed by one residual graph average, $\mathbf{E}=(1-\alpha)\mathbf{E}_0+\alpha\mathbf{P}\mathbf{E}_0$, where $\mathbf{P}$ is a row-normalized PPMI graph and $\alpha=0.3$. Terminal row normalization and per-dimension median/IQR scaling are then applied. On the Google analogy benchmark's family section, 272 of 506 questions are valid for every seed. Across five paired seeds, the complete pipeline raises accuracy from 19.41% to 30.74%, a gain of 11.32 percentage points with a nested-bootstrap 95% confidence interval of [6.93, 15.89]. Robust scaling alone contributes 3.24 points [1.25, 5.38], while graph averaging without robust scaling contributes 6.18 points [2.63, 9.92]. A separate 40-question general grid does not support a general improvement: the full pipeline changes accuracy by -6.00 points [-13.50, -0.50], and averaging without robust scaling changes it by -6.50 points [-14.50, -0.50]. The supported positive claim is therefore limited to the covered fairytales kinship analogy set; the results do not establish a generally effective embedding method.