Papers

Large language models (LLMs) have demonstrated strong capabilities in document key-value pair (KVP) extraction, yet controlled evaluations of their robustness to optical character recognition (OCR) output remain limited. This leaves an important gap in understanding their reliability in real-world OCR-to-LLM pipelines. Unlike end-to-end Vision-Language Models (VLMs), which jointly perform visual perception and semantic extraction, modular pipelines allow these stages and their errors to be isolated and audited. We introduce a controlled benchmark that distinguishes downstream LLM extraction behavior from upstream OCR degradation. It evaluates 136 experimental configurations and 17,688 document-level inferences generated with deterministic decoding across five instruction-tuned open-weight LLMs (2B-8B parameters), three datasets, and four text-quality conditions. The evaluation combines a full zero-shot comparison, targeted one- to three-shot experiments, and a sensitivity analysis of 40 configurations across 20 frozen demonstration sets. By separating Key Recall (annotated-field recovery) from Exact Match and Value F1 (exact and partial value recovery, respectively), we test whether OCR degradation affects field identification and value reproduction differently across models. Our findings challenge three practical assumptions: (1) clean-text performance reliably predicts real-world robustness, (2) model rankings remain consistent across annotation-derived Gold and OCR-derived text, and (3) additional few-shot demonstrations monotonically improve extraction accuracy. The observed model-ranking reversals and unstable few-shot gains expose important reliability risks under noisy document conditions. We release the benchmarking framework, dataset splits, and evaluation scripts to support reproducible research.

A decodable "empathy" direction is routinely read as a causal lever, conflating decodability, automated-metric control, and human-perceived change. We test this for two EPITOME-derived facets – Recognition (cognitive) and Resonance (affective) – in three instruction-tuned LLMs, scoring every intervention with two LLM judges and a discriminative EPITOME classifier, each gated by an emotional-vs-neutral positive control. The control passes for the affective facet across all automated instruments, but cognitive range is inconsistent across them. Both facets remain decodable after residualizing against a sentence-embedding-derived surface score, and steering can substantially rewrite the text. Yet adding the Resonance direction raises the affective score only partially – in Qwen by +0.29 (approximately 26% of the natural gap). A direct between-direction contrast confirms the shift is facet-specific in Qwen and Llama (not Gemma); we do not, however, establish a matching human-perceived change. Additive cognitive steering produces no measurable change, but a within-domain control shows the cognitive instrument is too coarse to resolve the differences such steering would produce – unmeasurable, not a clean null. By contrast, Gemma Recognition ablation lowers the classifier's cognitive score even after adjusting for response length. Detection does not imply reliable control under global interventions, and cognitive-empathy claims warrant an explicit measurement-sensitivity check.

Long-term memory is increasingly central to LLM agents, yet memory design remains a highly coupled architecture problem: what to encode, how to store it, how to retrieve it, and how to manage it can vary substantially across tasks and backbone models. We construct a discrete search space with 5 encoders, 5 stores, 6 retrievers, and 4 managers, and show that no single memory architecture consistently dominates: different tasks favor different module combinations, leading to substantial performance gaps. Motivated by this, we propose AutoMem, a text-gradient recursive self-improvement framework for task-adaptive memory architecture search. AutoMem optimizes over the factored space through two components: Experience-Guided Architecture Search, which proposes candidate architectures from historical search trajectories and accumulated reflections, and Failure-Guided Module Diagnosis, which localizes memory-related failures to specific modules and converts them into targeted textual feedback. Experiments on GAIA, WebWalkerQA, and xBench-DeepSearch across two LLM backbones show that AutoMem consistently discovers task-adaptive memory architectures that outperform the strongest human-designed memory baselines, improving accuracy by $2.8$ points on average across six benchmark-backbone settings. Further analysis shows that AutoMem achieves a favorable accuracy-efficiency trade-off, reducing token cost by $14.3\%$ over the strongest accuracy baselines under Qwen3.5-122B-A10B, while also finding stronger architectures than substantially larger random searches within only a few guided iterations.

Top-k Mixture-of-Experts (MoE) routing is discontinuous, so a deployment-motivated numerical disturbance – simulated 4-bit KV-cache quantization read by a protected BF16 gate – pushes tokens across decision boundaries and flips which experts fire. This paper proposes no new mitigation; it supplies a causal apparatus, empirical findings, and a detection-limit result. A four-run apparatus prices the route-mediated fraction (RMF) of quantization damage, a token-level attribution decomposes it by mechanism, and pre-registered probes carry the findings across three architectures. On OLMoE-1B-7B at 4-bit KV (pilot), about a third of the damage is routing-mediated: RMF ~ 0.31 (discovery 0.31 [0.20, 0.41]; process-replicated mean 0.313 +/- 0.020; pre-registered re-execution 0.231). The deployable router margin detects that a flip occurred (AUC 0.772) but cannot tell a harmful flip from a helpful one (at chance): among the tested local, inference-observable router statistics we find no predictor of a flip's loss sign above chance – an empirical benefit-detection barrier bounding selective repair restricted to this feature family. The signed-flip tax and sign-inseparability carry cross-model; the clean-reference remedy's payout is architecture-modulated; a controlled same-checkpoint flag-swap re-scopes the gate's normalization convention to a damage-magnitude moderator, not a route-recoverability mechanism. A real int4 KV kernel yields a fraction compatible with the fake-quant dose curve but underpowered (95% CI [-0.111, 0.394] includes zero) – ruling out gross disagreement, not an independent replication. Hypotheses, thresholds, and evaluations were pre-registered before measurement, with misses reported; a pre-registered held-out read replicates the partition and the near-cancelling tax out of sample, while the strict impossibility exclusion narrowly misses.

Most conferences rely on peer-review of submissions, but as generative AI makes it easier than ever to prepare submission materials, some conferences are seeing an overwhelming surge of submissions. We wanted to see if generative AI could help our conference's volunteer reviewers by pre-reviewing abstracts for certain criteria. The Bioinformatics Open Source Conference (BOSC) was well-positioned to experiment with this, as we already had a detailed rubric used by reviewers to evaluate submitted abstracts on multiple criteria, including openness (public availability of the code or other content associated with the project), valid open source license, and "runnability" (how easy it is to download, build, and run the project - an important measure of reusability). For BOSC 2026, we built bosc-pre-review, an agentic skill that assessed six review criteria, and Runabilly, which builds and tests each project in a disposable Docker container for safety. The AI only gathered evidence to present to the reviewers; humans made every decision regarding the acceptance of the abstracts. After the review period, we surveyed the reviewers to determine how useful they found the pre-review. Most of those who responded said they found it useful, but they preferred to check the AI's conclusions against their own, rather than accepting the AI results unquestioningly.

An agent harness is the external control layer that turns a base LLM into an executable agent by managing context, tools, orchestration, memory, decoding, and output handling. While harness design strongly affects agent behavior, most automatic improvement methods optimize narrower artifacts such as prompts, pipelines, or workflows, and deployed agents usually reuse a single global harness for all cases. We introduce MemoHarness, an adaptive harness optimization framework that learns from its own executions. MemoHarness decomposes the harness into six editable control dimensions, stores per-case diagnoses and distilled global patterns in a dual-layer experience bank, and adapts the learned harness to each test case using retrieved experience without test-time labels, feedback, or additional search. In our evaluation across shell-agent, code-generation, and analytical-reasoning benchmarks, MemoHarness improves over the fixed harnesses we compare against and shows selective transfer to unseen suites and base models. Its additional context can also remain cost-competitive when much of the retrieved experience is cacheable. These results provide evidence that execution experience is a practical substrate for building agent harnesses that are more adaptive than a single static configuration, while leaving broader claims about statistical robustness and component attribution to future work.

This position paper explores how Agentic AI and Model Context Protocol (MCP) can support power-grid studies in a Transmission System Operator (TSO) context. We focus on integrating Large Language Models with numerical simulation tools, structured workflows, and human supervision. We identify key industrial requirements for agent assisted grid studies and introduce pypowsybl-mcp, an MCP-based interface exposing selected capabilities of our simulation tool, pypowsybl to AI agents. This first step provides a testbed to study how agents can setup simulations, execute analyses, retrieve results, and interact with power-system simulators through standardized tool calls. We also discuss principles for human-in-the-loop, multi-agent workflows and outline an evaluation strategy combining technical metrics and practitioner feedback. The paper positions MCP-based tool integration as a step toward more interactive, auditable, and scalable grid-study environments.

Aligned language models refuse harmful requests, but a one-line prefill ("Sure, here is") strips the refusal. We ask where and how it fails. The harm representation stays intact: on the prompts the attack flips to compliance, a linear probe reads harm as high as on the refused ones (0.91-0.98), while behavioral refusal drops to chance. This holds across four models and three families (1.5-3.8B, and at 14B). Refusal is therefore a shallow, response-site computation. We localize it to an early window: a dose-matched position control shows the first half of the response suffices to break refusal, while the second half is nearly inert. Three causal probes converge on that window. Restoring the harm direction there partially re-engages refusal. Injecting the model's own refuse-state reverses the jailbreak (74%, held-out). And knocking out the early response's attention to the prefill, but not an equal attention mass elsewhere, selectively collapses the harmful continuation. A base-model control identifies the mechanism: the same knockout collapses the continuation prefill-specifically even in a non-safety-tuned base model (64% to 25% harmful content vs a matched control's 64%, replicated at 7B). So the prefill's grip is generic autoregressive conditioning, not safety-specific suppression, and "refusal restoration" is a model-dependent fallback. The dominant mechanism is passive. A small safety-specific attractor remains on top (logit-trace concentration 0.24 vs 0.03), whose active-vs-passive character we size but do not fully separate. No single direction or component is a clean handle either: the decision is decodable but distributed, and refusal tracks harm rather than scary surface. The consequence is structural: a monitor reading the untouched prompt-side representation is immune by construction, but only to response-site attacks. The mechanism is diffuse; the failure surface is local.

Unequal availability of human preference data across languages poses a significant challenge for aligning large language models in multilingual settings. To address the lack of sufficient data in low-resource language alignment, we propose a meta-learning framework for Reinforcement Learning from Human Feedback and Direct Preference Optimization. By leveraging preference data from other languages, our framework learns a transferable initialization that enables effective adaptation to a target language with minimal data. We provide theoretical guarantees for both the meta-reward modeling and meta-policy optimization settings, and empirically demonstrate the effectiveness of our approach on multilingual benchmarks. In an extremely low-resource setting with only 100 target-language preference samples, our approach achieves up to $28\%$ win-rate improvements over baseline methods, and consistently outperforms baselines across multiple target languages and model scales. Our approaches retain these advantages across different combinations of meta-training languages and varying linguistic distances from the target languages.

Retrieval in the SQL setting has largely been studied as the task of finding, within a large collection of SQL statements, the statement that answers a natural-language question. At scale, however, a more fundamental retrieval problem precedes generation: schema retrieval, identifying the tables and columns a question requires in a database that may contain thousands of them, far more than fit in a model's context. We argue that this step warrants first-class evaluation. To this end, we recast five text-to-SQL datasets (Spider, BIRD, BEAVER, and two LiveSQLBench variants) as retrieval tasks at both table and column granularity, covering realistic and enterprise-scale schemas under two document representations, and we show that off-the-shelf text and code embedders transfer poorly to this setting. We then propose corpus-adaptive fine-tuning: natural-language queries are synthesized directly from the target schema corpus, granularity-aware hard negatives are mined, and a 305M-parameter embedder is fine-tuned contrastively. This procedure raises average recall@10 from 60.4 to 75.6 (nDCG@10 from 51.9 to 68.0), making the 305M model the strongest retriever under one billion parameters and competitive with state-of-the-art embedders of 4-8B parameters, more than an order of magnitude larger. The same recipe improves an 8B state-of-the-art embedder from 77.8 to 78.4 recall@10, matching the best result on the benchmark and indicating that the adaptation is backbone-agnostic. Leave-one-corpus-out experiments and a leakage audit show that these gains reflect a transferable schema-retrieval ability rather than memorization of the evaluation data. Our results establish schema linking as a standalone retrieval task and lightweight, label-free corpus adaptation as a practical route to deploying it at enterprise scale.

Teams measuring whether large language models (LLMs) recommend a brand face a reproducibility problem: ask the same question twice and the answer moves. Practice resamples each prompt a few times (commonly five) and averages, treating within-prompt resampling as the source of the noise. But a measured brand score moves for at least four separable reasons: within-prompt resampling, prompt paraphrase, model identity, and query language. We specify a crossed random-effects (generalizability-theory) decomposition that partitions the total variance of a response-level brand outcome into these four sources, and embed the components in a decision-study allocation that returns how many repeats, paraphrases, models, and languages to buy for a target reliability. We apply it to a fully crossed corpus of 12,933 LLM responses on 20 Central and Eastern European brands, 8 languages, and 3 models (GPT-5.2 and Gemini 3 Flash in parametric mode, Perplexity in grounded retrieval), with a stability subset of 1,435 cells resampled about five times. The outcome is per-response multilingual sentiment polarity. Query language is the largest systematic facet (26.5% of the variance of one response) against 1.5% for brand identity (ICC 0.0146), so a single AI answer carries almost no brand-discriminating signal. Once a cell term isolates pure resampling, resampling is 34.8% of variance and the brand-in-context interaction 29.6%; brand-by-language is 8.6% (a bilingual penalty) while brand-by-model and brand-by-prompt are near zero. Per unit of query budget, adding languages and models reduces relative-error variance far more than adding repeats: a repeat past the fifth reduces it by only 0.0003. Brand-ranking reliability stays low, near 0.01 for a single answer and about 0.36 at the full crossed design, so reliability is bought by spreading across languages and models, not by repeating one prompt.

Autoformalization translates informal natural language into formal, machine-verifiable languages. While most work focuses on individual statements, real formalization efforts are inherently theory-level: they require an entire web of axioms, definitions, and lemmas before target theorems can even be stated. In this position paper, we argue for theory-level autoformalization: formalizing complete theories, including all their inter-dependencies, as structured libraries. We examine the significance of this shift, address alternative views, identify open challenges, and propose three promising paths forward. Our survey of autoformalization is available at https://github.com/marcusm117/Awesome-Autoformalization.

Policy documents shape governance outcomes, but their reasoning is often implicit. Participatory commitments and managerial control routinely coexist in the same text, and the tensions between them are rarely stated directly. Existing computational approaches to policy discourse cannot express the frame-mediated relations that drive these tensions, where one argument narrows or instrumentalizes another rather than rejecting it. End-to-end summarization by large language models produces fluent text but offers little structure that domain experts can inspect or contest. We present Apaf, a hybrid LLM–symbolic pipeline that operationalizes critical discourse analysis as a quantitative bipolar argumentation framework over policy text. Arguments are first classified into deliberative or managerial frames. Four frame-mediated relation subtypes (agency reduction, agenda shift, instrumental support, and normative support) are then produced by deterministic rules over LLM-extracted features. We release a novel dataset of 100 sub-documents of disaster-risk-reduction policy from the USA, UK, Canada, and Australia, and show that the resulting argument graphs are accurate, interpretable, and stable across jurisdictions.

The evaluation of mathematical reasoning in large language models (LLMs) has predominantly focused on high-resource languages like English. This has created a significant barrier to the equitable development and deployment of AI in linguistically diverse regions such as Bangladesh, where over 230 million people speak Bengali. Despite this global significance, there has been minimal prior work on mathematical reasoning in Bengali and no existing research that systematically benchmarks a perturbated Bengali mathematical dataset, leaving a critical void in assessing model robustness and true comprehension beyond pattern recognition. This study addresses this gap by introducing GSM-Plus-BN, a novel perturbated Bengali mathematical dataset derived from the English GSM-Plus benchmark and verified by human translators. We evaluate six open-source LLMs Qwen3-32B, Llama-3.1-8B-Instant, Llama-3.3-70B-Versatile, Llama-4-Scout-17B-16E-Instruct, GPT-OSS-120B, and GPT-OSS-20B using a benchmark of 9,000 evaluation samples comprising 1,000 seed questions and 8,000 perturbed variants under both Standard Prompting and Chain-of-Thought (CoT) Prompting. Experimental results show that GPT-OSS-20B achieves the highest seed question accuracy of 96.08% under Standard Prompting, while larger models such as Llama-3.3-70B and GPT-OSS-120B demonstrate superior robustness across perturbation types. Furthermore, CoT prompting substantially improves reasoning for most models compared to Standard Prompting, yet a notable performance gap persists across all models relative to their English benchmarks, underscoring the inherent difficulty of perturbed Bengali text. This research makes a foundational contribution by providing GSM-PLUS-BN as a new resource and baseline for future Bengali mathematical reasoning research.

Attention-based KV cache eviction (H2O and its descendants) compresses the memory-constrained state of a long-context model by ranking tokens on accumulated attention mass, treated here as signal energy, and keeping the heaviest. On schema-dense input streams such as nested JSON, this score acts as a non-stationary filter that disproportionately retains noise: a non-content sink role (delimiters or whitespace) carries an order of magnitude more energy than any content role, and structural KEY tokens are over-retained at roughly 1.8x the rate of the answer-carrying VALUE tokens, collapsing exact-match accuracy from 88% to 0% at a 5% budget as the signal-to-noise ratio of the retained state degrades. A counterfactual experiment establishes that suppressing KEY tokens is the best deployable filter. Our retraining-free, role-conditional allocation over SnapKV's windowed score, governed by a single tuned hyperparameter, closes 63-98% of the H2O gap at sub-20% budgets and, at higher budgets, modestly matches or exceeds full-cache accuracy – a small, seed-sensitive denoising effect (borderline significant at B=0.50; not distinguishable from zero at B=0.30 over four seeds). A 15 MB linear role probe supplies these labels at negligible inference cost, though matching parser-level downstream accuracy remains open.

We present RAGthoven, our system for SemEval-2026 Task 1 (MWAHAHA), Subtask A (multilingual constrained humor generation in English, Spanish, and Chinese). RAGthoven decomposes creative text generation into a multi-stage large language model (LLM) pipeline (Planner, Best-of-N Writer, Reflector for self-critique, LLM-as-a-judge Judge) grounded in computational humor theory (Benign Violation Theory, Script-based Semantic Theory of Humor) and refined across ten experiments. In our final configuration, we augment the Planner with retrieval-augmented generation (RAG) from a curated joke corpus, seeding generation with diverse joke mechanisms. We also evaluate two agentic variants – ReAct-style sequential tool-calling (Exp09) and autonomous multi-branch orchestration (Exp10) – that expose the same four stages with a deterministic ConstraintAudit checker. Across four frontier models on a held-out 12-instance English sample, neither agentic variant produced outputs we judged superior to the non-agentic pipeline despite substantially higher tool-call budgets. RAGthoven shares Rank 1 with the Gemini 2.5 Flash baseline in all three languages, with overlapping organizer-reported confidence intervals. In Spanish, it leads the baseline by 42 raw Elo points (1182 vs. 1140), while in English (1045 vs. 1081) and Chinese (1045 vs. 1053) the baseline holds the higher raw rating within the same statistical tie. Together, these results suggest language-dependent diminishing returns from elaborate multi-stage prompt engineering and agentic scaffolding once a strong frontier model is in the loop.

Sign language is a primary communication channel for millions of Deaf and hard-of-hearing people, yet text-to-signer video generation remains costly because video diffusion models are expensive to train and evaluate. This paper presents Text2Sign, a text-conditioned diffusion model for short sign-language clips that runs on a single NVIDIA L4 GPU. It combines a frozen vision-language text encoder with a 3D encoder-decoder and factorized spatiotemporal attention to reduce the cost of full-video attention while preserving motion coherence. We compare convolution-only and transformer-style backbones, frozen pretrained and task-specific text encoders, and factorized versus full attention. On a signer-disjoint How2Sign split, the best short-run ablation reaches a validation loss of 0.0648, while a longer-run checkpoint reaches 0.00999. On a compact evaluation slice, the latter achieves an SSIM of $0.2403 \pm 0.0238$, a PSNR of $15.11 \pm 0.42$ dB, and temporal consistency of $1.0000 \pm 0.0000$ using 8-step DDIM sampling with a guidance scale of 5.0. It generates a 32-frame, $64 \times 64$ clip in 12.60 seconds, or 2.54 frames per second, with peak inference memory of 3.12 GB. A held-out denoising audit shows only weak prompt sensitivity: removing text increases late-timestep loss from 0.9875 to 0.9891, while shuffled prompts perform similarly to correct prompts. Frozen text conditioning therefore improves short-budget validation loss, but prompt-specific separation remains limited. The system is restricted to low-resolution, short clips and lacks expert linguistic evaluation, so it should be viewed as a single-GPU research baseline rather than a complete sign-language production system. Code is available at https://github.com/xiaruize0911/text2sign.

What a language model will and will not do is largely set during post-training, but which behaviors it expresses, hides, or resists is not revealed by prompting alone. Persona vectors, behavioral directions in activation space, can probe this organization, but prior work covers only a handful of traits. We present the first systematic application of persona vectors at this scale, compiling a 53-trait inventory across four behaviorally distinct domains and labeling every trait in two open-weight models as natural (expressed at baseline), steerable latent but amplifiable, or intractable (resistant to standard extraction). Both models default to helpful, task-oriented behavior: all nine agentic traits are natural, and their default clinician behavior matches a board-certified psychologist's independent desirability judgments on 16 of 17 traits. Steering produces its largest gains on traits these defaults exclude: hyperbole, hallucination, and sycophancy. The same asymmetry holds across all 171 generic-trait pairs: two steerable traits can collapse the composition, but pairs involving a default never do. Where standard extraction fails on a trait like "evil," a vector transferred from a fine-tuned variant still recovers it, with the residual refusals appearing inside the model's chain-of-thought. Persona vectors are most informative not as a set of controls but as a probe of behavioral organization.

Simultaneous speech translation (SimulST) requires incremental translation under strict latency constraints, yet remains challenging for decoder-only LLM systems due to limited context and cross-lingual reordering. Recent approaches often introduce architectural changes or explicit read/write policies to control output timing, which can be brittle in conversational speech where segmentation boundaries are ambiguous. We present a simple data-driven alternative: fixed-length chunks for cumulative streaming decoding with a rewind-based committed prefix, and teacher-labeled prefix-to-prefix (P2P) targets with bounded waiting for fine-tuning, yielding CSSEL-P2P, where CSSEL is our proposed chunked streaming speech encoder LLM. In our in-house conversational speech evaluation, CSSEL-P2P improves streaming quality by +1.54 COMETKiwi over the CSSEL streaming baseline at comparable latency (+0.15s Average Lagging), suggesting effective SimulST without architectural changes via P2P supervision.

Structured pruning is a hardware-friendly way to compress LLMs, but it is mostly validated on multiple-choice recognition tasks, while the same compressed checkpoints can collapse on the free-form generation that deployment actually requires. Two observations trace this gap. First, greedy pass@$1$ nearly vanishes after compression, yet pass@$k$ recovers substantially under repeated sampling: useful generations are demoted, not erased. Second, the recoverable regime fails mainly through suffix repetition. Recovery should therefore train on the compressed model's own on-policy states with dense token-level supervision, which On-Policy Distillation (OPD) provides by reusing the pre-compression model as a frozen teacher. However, long on-policy rollouts spend early recovery budget on low-information repetitive suffixes, delaying loss descent. To mitigate this waste, we propose \shortopd, a short-to-long OPD schedule that detects teacher-confirmed repetitive suffixes, treats the surviving prefix as each rollout's effective length, and allocates future rollout budgets to the effective lengths the policy can currently use. Across math, code, and open-ended generation, \shortopd\ raises the compressed model's score to about $9\times$ its unrecovered value and $1.6$–$4.4\times$ standard recovery recipes (SFT w/o KD, KD, and SeqKD), and it matches a fixed $8192$-token rollout horizon within two points using a quarter of the training time ($8.5$ vs.\ $35.9$ hours) and $71\%$ fewer rollout tokens. We hope this recipe helps move structured pruning beyond marginal gains on perplexity and multiple-choice benchmarks, a step closer to deployment-ready generation quality.

Self-improving autonomous agents are moving from research prototypes to deployed systems. The primary goal is controllable evolution, or adaptation, from experience with minimal or even no human input. This survey frames modern self-improving agents as adaptive systems that convert experience into accumulated capability gains. We offer a system-level framework that represents a modern agent as a configuration coupling a foundation model with an operational scaffold of prompts, memory, tools, and control logic. Within this framework, self-improvement is formalized as a self-induced update operator that obtains and commits updates to model parameters or scaffold components. We organize prior work by update target and by the signals that drive change, then review applications and discuss evaluation, before closing with open problems and future directions. For convenience, we track technical updates on https://github.com/selfimproving-agent/awesome-Self-Improving-Agents.

Large language model (LLM) agents increasingly automate multi-step engineering and informatics workflows, yet they rarely ask how much effort a task actually requires. They often follow a maximum-context-first strategy–re-reading files and dependencies they have already seen–turning a one-line edit into a small code-base audit. We argue the missing capability is task-aware execution-scope estimation: judging a task's difficulty, the information it truly needs, and the shortest reliable path before committing budget. We formalize minimum-sufficient execution and the Agent Cognitive Redundancy Ratio (ACRR), and propose E3 (Estimate, Execute, Expand): the agent estimates an initial operating point, executes a minimum viable path, and expands scope only when verification fails. On MSE-Bench–a deterministic benchmark of 121 edits in a capability-controlled simulator–E3 matches the strongest baseline's 100% success while cutting cost by 85%, tokens by 91%, and inspected files by 92%, and further beats a strong adaptive retrieval baseline by 16%; the gains survive held-out instruction wording and essentially every cost weighting. A companion real-model harness (LLM-Case) corroborates the effect on a live gpt-4o agent editing a real open-source library, with every candidate patch graded by actually running the project's real pytest suite against a measured oracle: the over-reading is milder but real, and E3 is the leanest and fastest policy at comparable task success–its one shortfall a provider rate-limit, not a wrong edit. We frame this as a controlled probe of execution redundancy, not a measurement of any deployed agent, and position task-aware execution as a step toward engineering-grounded AI (EGAI)–agents whose effort is anchored in the engineering reality of the task. We release the framework and benchmark.

Large Language Model (LLM) agents have moved beyond generating responses to executing multi-step tasks by calling tools, observing the results, and iteratively deciding the next action. Most agent systems run on desktops or servers, which support tool use and task automation. Mobile devices are also important agent environments because they are widely accessible and contain users' data, sensors, and daily-use applications. Existing mobile agents mainly operate smartphones through graphical user interface (GUI) actions such as tapping, swiping, and typing, which often form long, interface-dependent sequences, cannot directly access device capabilities, and make execution boundaries difficult to define. We present PalmClaw, an open-source agent framework that runs natively on mobile phones and manages the sessions, memory, skills, tools, and agent loop directly on the device. PalmClaw exposes device capabilities as device tools with explicit arguments, structured results, and clearly defined execution boundaries. This design enables agents to use mobile capabilities directly while keeping each action explicit and controlled. Experiments show an 11.5% relative improvement in task success and a 94.9% reduction in completion time over the strongest baseline, with lower setup burden and traces illustrating how execution boundaries are applied. Code is available at https://github.com/ModalityDance/PalmClaw.

As large language models (LLMs) grow more capable, they are increasingly deployed in context-rich settings where task inputs are often accompanied by long, partially irrelevant context. In a controlled setting, we find that state-of-the-art models often appear robust to task-irrelevant context at the aggregate level: prepending it to benchmark questions causes little change in overall accuracy. This aggregate stability, however, masks significant per-example instability. Even semantically meaningless pseudo-words, formed by randomly combining characters, can markedly shift model predictions on a small fraction of examples, degrading performance on some while improving it on others. This two-sided effect holds consistently across a wide range of models and datasets, yet the affected examples are largely model-specific. We further show that this instability is modulated by context type, context length, test-time compute, and model development stage. Together, our findings reveal context-induced tail risks concealed by aggregate accuracy, motivating per-example reliability evaluation of language models.

Long-term memory has become a foundational capability for LLM-based agents that accompany users across extended, multi-session interactions. Existing benchmarks, however, evaluate such memory almost exclusively through downstream question answering, scoring only the correctness of a final answer. This black-box formulation conflates the heterogeneous causes of memory failure, such as missing the introduction of a relevant fact, binding an operation to the wrong target, or relying on stale values after a correction. As a result, it can credit correct answers despite their reliance on inconsistent or unsafe memory states. In this paper, we argue that, in dynamic long-horizon interactions, memory is not a static collection of facts but a lifecycle of explicit operations, including remembering, forgetting, updating, reflecting, and their compositions. We introduce MemOps, a benchmark that reformulates conversational memory as a sequence of lifecycle operations and represents each memory event with a structured trace specifying its trigger, target, scope, state transition, and supporting evidence. A controllable generation pipeline embeds these operations into long, task-oriented conversations and produces gold operation traces together with six categories of operation-level probes, evaluated under both adjacent-evidence and long-context settings. Across long-context, retrieval-based, parametric and managed-memory systems, MemOps disentangles failure modes that final-answer accuracy alone conceals, revealing that current systems remain far from uniformly reliable. For instance, session-level retrieval outperforms turn-level retrieval, and long-context models remain notably weak at reconstructing ordered memory-state trajectories. These results move long-term memory evaluation from final-answer scoring toward interpretable, operation-level diagnosis.

LLM judges are increasingly being used to evaluate open-ended model responses, often in no-reference settings where a ground-truth answer is unavailable. However, can they reliably assess in such evaluation setups? We explore this question in this paper through a two stage pipeline with a) calibration experiments that assess the judge model's knowledge of the task it is evaluating, and b) sensitivity experiments that assess how the judge model's performance is impacted by the presence and positioning of the reference answer in the prompt. Across experiments covering three languages, we show that the judge models we evaluated tend to over-credit incorrect answers in the absence of a reference answer, and adding reference answer information to the prompt flips the judge model's correct/incorrect decisions by as much as 85% in some experimental settings. Comparison with a subset of human annotations shows that these reference-driven changes generally align with human judgments. Our results emphasize the need for calibrating the LLM judges with a sample with reference-aware evaluation before using them in reference-free setups reliably, and our methodology provides a blueprint for researchers and practitioners in doing such calibration of LLM judges for other tasks.

Patients seeking medical information often ask questions that embed incorrect assumptions or misconceptions. In such cases, safe medical communication requires not only answering the question, but identifying and correcting the underlying false belief. These interactions naturally unfold over multiple turns, a pattern now mirrored in interactions with LLMs. Yet current evaluation frameworks do not capture model behavior in these settings, where misconceptions can emerge, persist, or evolve over the course of a conversation. Whether LLMs can reliably correct such misconceptions over time remains largely unexamined. To study this, we introduce ThReadMed-QA, a multi-turn medical dialogue dataset of 2,437 patient-physician conversation threads comprising 8,204 question-answer pairs, derived from real patient interactions on AskDocs. This dataset enables systematic evaluation of whether models can detect and correct misconceptions under a multi-turn context. We evaluate five LLMs using a rubric-based LLM-as-a-Judge framework that scores responses based on their ability to identify and correct misconceptions. Our experiments reveal a consistent pattern: even frontier models that can address misconceptions in a single interaction degrade substantially over subsequent turns. GPT-5 and Claude-Haiku correct these false presuppositions around 85% on initial questions but drop to roughly 50% within two follow-ups. An oracle analysis replacing prior model outputs with physician responses shows that much of the degradation is driven by error propagation, while performance remains imperfect even under correct context. Even when models tend to correct misconceptions initially, their performance degrades substantially over later turns, leading to inconsistent and potentially unsafe guidance in patient-facing settings and highlighting the need for evaluation frameworks that capture multi-turn behavior.

Rubric-based evaluation is a promising approach for assessing open-ended outputs from LLM-based research agents, particularly in paper reproduction, where direct paper-to-repository comparison is prone to hallucination. However, constructing paper-specific rubrics requires substantial expert effort, limiting the scalability of benchmarks such as PaperBench. In this work, we present, to our knowledge, the first systematic meta-evaluation of LLM-generated rubrics for paper reproduction. We reformulate rubrics into a checklist-style format and evaluate four generation settings across two backbone models. We meta-evaluate generated rubrics intrinsically by semantic similarity and extrinsically by score alignment with ground-truth rubrics. Our results show that the augmented settings substantially improves downstream evaluation alignment, with the strongest setting approaching the human baseline, while intrinsic gains are more modest. Further analyses reveal that LLM-generated rubrics are often overly fine-grained, biased toward high scores, and less adaptive to paper domains, highlighting both the affordances and limitations.

Language models encode substantial factual knowledge in their parameters, which can lead to unreliable behavior when this knowledge is outdated, incomplete, or misaligned with the provided context. In this work, we study whether modifying the pretraining signal can systematically shift models away from parametric recall and toward evidence-grounded reasoning. We introduce Knowledge–”Less” Language Models (KLLMs), a fundamentally different epistemic training paradigm for LLMs, which are pretrained on corpora in which named entities are anonymized, thereby removing a primary channel for entity-linked factual supervision. This intervention substantially reduces closed-book factual recall, while often improving performance on tasks where relevant information is provided as context. Across multiple model scales, KLLMs consistently outperform matched baselines on contextual question answering, fact verification, and hallucination detection benchmarks. Crucially, in retrieval-grounded settings with imperfect evidence, KLLMs show improved robustness and achieve up to 20–25% relative gains over standard language models. They further exhibit better calibration, with improved ECE, Brier score, and AUROC, as well as more reliable abstention behavior. Our results demonstrate that suppressing entity-linked supervision during pretraining induces a shift in epistemic behavior: KLLMs rely less on parametric knowledge and more on external evidence, leading to improved reliability under realistic conditions. This suggests that pretraining-time control over knowledge acquisition can complement retrieval-augmented and tool-based systems by providing a more evidence-sensitive base model.

Diffusion large language models (dLLMs) offer a theoretical advantage in parallel generation over standard autoregressive models. However, parallel generation alone does not guarantee practical speedups. Realizing this efficiency requires specialized inference mechanisms, such as diffusion-aware caching and reuse. Consequently, as inference efficiency becomes a prerequisite for practical deployment, recent research has actively explored acceleration techniques across algorithms, architectures, and systems. However, rigorous comparisons remain difficult, as end-to-end latency stems from intricate trade-offs between algorithmic, architectural, and system-level factors that are often conflated in existing benchmarks. In this survey, we introduce a unified latency decomposition framework for dLLMs to disentangle these factors and analyze their impact on inference speed in real deployments. Guided by this framework, we categorize acceleration techniques along three axes covering algorithmic innovations, architectural and system optimizations, and inference-time scaling. Finally, we provide guidelines for reproducible benchmarking and highlight open challenges for realizing the full potential of parallel generation.