SEARCH · Searcharxiv

Search Searcharxiv

Search indexed arXiv papers on artificial intelligence, large language models, computer vision and robotics. Read source abstracts and follow links to arXiv.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 73 records · Page 4Linked to original sources

arXiv2026

Stratified Consistency Distillation for Natural Language Formalization

Neurosymbolic reasoning has shown promising success in addressing complex reasoning tasks by combining large language models (LLMs) and symbolic solvers. While this approach shows promise, a fundamental challenge remains: improving the accuracy of translations from natural language to logical formulas. Current methods predominantly rely on prompt engineering, which is difficult to scale across different domains and input formats. Drawing inspiration from the success of fine-tuning in other model adaptation and alignment applications, we propose a fine-tuning-based Stratified Consistency Distillation approach: (1) We generate K logical translations per input using a frontier LLM and cluster them by semantic equivalence (2) Based on the entropy level, we apply majority voting (low entropy), LLM-as-a-Judge (medium entropy), or unification/abstention (high entropy), and (3) fine-tune a smaller model using the selected pseudo-labels. Our experiments show significant and consistent improvements in both Pass@K and our novel Equivalent Logical Similarity metrics, demonstrating the potential of advancing logical translation through consistency distillation.

cs.CL↗

arXiv2026

Will the User Ever Know? Covert Indirect Prompt Injection Attacks on Tool-Using LLM Agents

As LLM agents take real-world actions through tools, indirect prompt injection (IPI) has emerged as a serious threat. The standard metric, Attack Success Rate (ASR), counts whether an injection succeeds but ignores what the user notices in the agent's final response. Looking at successful injection traces, we find two distinct outcomes: the agent executes the injection while returning an otherwise normal response, or reports the injected action in its final response, giving the user a chance to notice. We call these covert and overt successes. From the user's perspective, we decompose ASR into the Covert Success Rate (CSR), counting successes leaving no trace in the final response, and the Overt Success Rate (OSR), counting successes the user can detect. To understand what drives the gap, we analyze successful trajectories and find that the agent's behavior after the injection separates covert from overt: covert traces hand control back to the user task before ending, while overt traces end at the attack itself. This split follows from the ReAct format, where the final response summarizes the most recent action. Building on this observation, we propose ICoA (Induced Covert Attack), an IPI attack designed to induce covert outcomes by steering the agent back to the user task after executing the injection. Across four target models on AgentDojo, ICoA achieves the highest CSR, with gains of 3.79-12.01 percentage points over the strongest baseline.

cs.AI↗

arXiv2026

trajectory-judge: What Outcome-Only LLM Judges Miss on Agent Trajectories

A direct test of an LLM judge of agent trajectories injects faults into correct runs and reports recall, per fault type or by whether the fault broke the environment outcome (loud) or not (silent). Such recall can credit a judge with detection it does not have; paired discrimination, its flag rate on the faults minus its rate on the clean runs they came from, exposes this. Our testbed, a deterministic support desk with a scripted oracle and a one-step fault injector, labels all 400 trajectories exactly. A 14B judge shown only the request and final reply scores 34% to 76% recall on four fault types that leave the reply unchanged. There its input is the clean run's, so its paired discrimination is zero and that recall is its flag rate on clean runs. Splitting by outcome survival does not fix this: its loud recall of 84% is a paired +0.393 and its silent recall of 45% a paired +0.048, all from the two fault types that change the reply. Told to check each step, the same model flags every fault of those four types and 0 of 100 clean runs (95% CI up to 3.6%). It does not reliably check the reply: of four invented promises it flags one every time and the other three once in 42 faults. Shown every step but asked only about the reply, it still reaches a paired +0.69 on reply-unchanged faults, against +1.00 when told to check each step. We recommend reporting paired discrimination against clean parents, split by whether the fault reaches the judge's input and by outcome survival, and release the testbed, raw verdicts and analysis pipeline.

cs.CL↗

arXiv2026

Intelligence per Watt: Measuring Intelligence Efficiency of Local AI

Large language model (LLM) queries are predominantly processed by frontier models in centralized cloud infrastructure. Demand growth strains this paradigm faster than providers can scale. Two advances create an opportunity to rethink it: small, local LMs (<=20B active parameters) now achieve competitive performance to frontier models on many tasks, and local accelerators (e.g., Apple M4 Max) can host these models at interactive latencies. This raises the question: can local inference viably redistribute demand from centralized infrastructure? This requires measuring both whether local LMs can accurately answer real-world queries and whether they can do so efficiently on power-constrained devices (e.g., laptops). We propose intelligence per watt (IPW), task accuracy per unit of power, as a unified metric for the capability and efficiency of local inference across model-accelerator configurations. We evaluate 20+ state-of-the-art local LMs, 8 hardware accelerators (local and cloud), and 1M real-world single-turn chat and reasoning queries. For each query, we measure accuracy (local LM win rate against frontier models), energy, latency, and power. We find three key results. First, local LMs successfully answer 88.7% of these queries, with accuracy varying by domain. Second, longitudinal analysis from 2023-2025 shows IPW improved 5.3x, driven by both algorithmic and accelerator advances, with locally-serviceable query coverage rising from 23.2% to 71.3%. Third, local accelerators achieve at least 1.4x lower IPW than cloud accelerators running identical models, revealing significant headroom for local accelerator optimization. These findings demonstrate that local inference can meaningfully redistribute demand from centralized infrastructure for a substantial subset of queries, with IPW serving as the critical metric for tracking this transition.

cs.DC↗

arXiv2026

Aligning Language Model Benchmarks with Pairwise Preferences

Language model benchmarks are pervasive and computationally-efficient proxies for real-world downstream performance. However, many recent works find that benchmarks often fail to predict downstream utility. While some works have begun diagnosing sources of misalignment, there remain no ways to systematically update benchmarks to align their scores with downstream usage. Towards bridging this gap, we introduce and study \textit{benchmark alignment}, where we use information about downstream model performance to automatically update benchmarks, specifically aiming to update static benchmarks so they generalizably rank models according to new pairwise preferences. Our experiments involving 4576 language models and 6 benchmarks show that reweighting benchmark items can successfully rank unseen models, even generalizing across model scales in most cases. And while naive alignment unsurprisingly requires large numbers of models and benchmark questions, an oracle experiment suggests this could be reduced to as few as 20 well-chosen models. Overall, our work takes a step towards efficiently aligning benchmark development with downstream tasks.\footnote{All of our code, models, and data are publicly-available.

cs.AI↗

arXiv2026

One Success Isn't Reliability: Thinkingbox, a Sandbox and Benchmark for Agents in Stateful Business Workflows

Recent agent benchmarks increasingly ground evaluation in executable environments, from code repair to web navigation, app APIs, and function calling. Yet completing consequential work beyond code requires more than producing a plausible response or valid tool call: agents must gather missing information over multiple turns, follow domain policies, coordinate dependent tools, and realize the correct persistent state transition without collateral effects. In this paper, we introduce Thinkingbox, a sandbox for tool-agent-user interaction that provides isolated MCP-compatible tool sessions, complete execution traces, and outcome evaluation over terminal backend state. Built on this sandbox, Thinkingbox-bench contains 507 policy-conditioned workflows across business scenarios, including retail, hospitality, auto insurance, neobank internal IT, and consulting IT/HR support. Each attempt is evaluated by task-specific executable checks that accept valid trajectories while rejecting wrong, missing, or extra effects; designated tasks additionally check required properties of the final response. Our experiments reveal that even the strongest proprietary and open-weight models show steep reliability drops: Claude Opus 5 falls from 66.50% pass@1 to 47.53% pass^20, and Kimi-K3 from 57.37% pass@1 to 17.60% pass^20. Moreover, many failed trials terminate cleanly after valid state-changing actions, so response- or tool-call-level signals poorly proxy end-to-end completion. Thinkingbox-bench reveals a large gap between occasionally finding a successful trajectory and reliably completing stateful business tasks. We release both Thinkingbox (https://github.com/microsoft/thinkingbox) and Thinkingbox-bench (https://github.com/microsoft/thinkingbox-data).

cs.CL↗

arXiv2026

SE-GoS: Self-Evolving Graph-of-Skills for Skill Library at Scale

LLM agents use large libraries of reusable skills. At thousands of skill entries, retrieval becomes the bottleneck. Graph-of-Skills (GoS) retrieves dependency-aware bundles from a typed skill graph, and SkillDAG shows that such a graph can accumulate execution-backed structure online. Neither asks whether execution traces can be distilled into a better retrieval graph that generalizes to unseen tasks. We present \textbf{Self-Evolving Graph-of-Skills (SE-GoS)}, which treats the retrieval graph as an index rather than a learned representation: the graph is maintained from execution traces while the retrieval pipeline, the skill library, and the model stay fixed. SE-GoS applies three updates: (1) \textbf{topology}, which induces relations from execution evidence and retracts an avoid edge only after repeated successful co-use; (2) \textbf{edge-weight}, which softly attenuates unsupported semantic edges and reinforces incoming edges to used skills; and (3) \textbf{node-description}, which updates retrieval-facing descriptions stored on graph nodes ranked too low. On SkillsBench, one evolution round lifts average reward from 52.4\% to 59.4\%, above full-library loading, vector retrieval, static GoS, and SkillDAG, and this ordering repeats on all three backbones. Retrieval over the evolved graph spends about two-thirds of the input tokens that loading the full library costs. Repeating the round does not help. The same graph improves a held-out split it never saw from 52.9\% to 58.3\%, so what it accumulates transfers rather than memorizes traces. Skill graphs can therefore be improved from execution experience without model training, retrieval-algorithm changes, skill-content modifications, or a model judging which skills are related.

cs.AI↗

arXiv2026

Distributional Validity of a Korean Synthetic Persona Panel: Evidence From the Korea Media Panel Survey

Large language model (LLM) personas are proposed as survey respondents, yet validation outside English-speaking contexts is scarce. We evaluate how well a Korean synthetic persona panel used to condition Gemini 3.5 Flash and EXAONE reproduces digital and artificial intelligence (AI) service-use distributions of the Korea Media Panel Survey. About 8,000 personas per model answered eight service-use items and eight attitudinal constructs; responses were compared with weighted survey estimates. The overall mean absolute error (MAE) was 14-19 percentage points (pp), with binary item-mean correlations of 0.70-0.91 across waves. Segment error across five axes was 14-18 pp, with between-group signed-error ranges of 49.6/34.7 pp (Gemini/EXAONE; 39.5/31.2 without the non-comparable teen cells). Errors were model-specific: an age stereotype (Gemini) versus an acquiescence-consistent level bias (EXAONE). Generative-AI overestimation was consistent with temporal misalignment; short-form underestimation was framing-sensitive and persisted under randomized order (both shown for Gemini). Post-hoc holdout calibration on 30% of the real data, with the correction form selected inside the calibration set, cut cell MAE from 18.3/15.1 to 4.9/4.4 pp, yet direct estimation from that subsample was more accurate than the calibrated panel (3.6 pp), a synthetic-informed shrinkage estimator beat its real-only counterpart by at most 0.7 pp, and the correction did not transfer competitively across waves. The calibrated panel kept an advantage only below roughly 250-860 real responses (at most 2.3 pp over a real-only shrinkage estimator) or, for one model, on unobserved segments. In this setting, synthetic panels are not survey substitutes; their value is diagnostic.

cs.CY↗

arXiv2026

LLM Abstention Can Be a Prompt Artifact, in Addition to Genuine Uncertainty

Large Language Models (LLMs) are increasingly trained to abstain from answering questions they are unsure about. However, this ability is often misapplied: in real-world applications, user prompts sometimes contain elements of uncertainty, which lead LLMs to abstain even on problems they are capable of solving. We argue that LLM abstention is not only an expression of genuine uncertainty; it can also be an artifact largely shaped by prompts. We name this phenomenon *Abstention Inflation*. We add "Unknown" as an extra option for LLMs to choose from; experiments show serious accuracy drops on True/False Questions (TFQs). Replacing "Unknown" with an unrelated random word produces a similar effect. We argue that LLMs are trained to imitate the surface pattern of abstention, rather than to express genuine uncertainty. Based on ten experimental settings, we support four claims that form a progressive argument: **(C1)** *Abstention Inflation* can be triggered by the presence of an extra option, not by genuine uncertainty; **(C2)** it makes the models deny they can answer, even when they can; **(C3)** it is a later-layer output override, as the reasoning traces and mid-layer representations preserve correct answers; **(C4)** it is not stochastic noise: it results from various factors, emerges through instruction tuning, is boosted by problems' higher difficulty, and can be mitigated at larger model sizes.

cs.CL↗

arXiv2026

Correct Prediction, Wrong Steps? Consensus Reasoning Knowledge Graph for Robust Chain-of-Thought Synthesis

Large language models (LLMs) have become increasingly used for various tasks, often coupled with Chain-of-Thought (CoT) prompting to boost accuracy. Recent work has shown that high label-prediction accuracy does not guarantee correct intermediate reasoning, and the causes of *reasoning flaws* vary from sample to sample, yet existing remedies either focus on a single domain or assume that one flaw type applies uniformly across samples. A simple mitigation method is to provide the model with the correct answer, but we show that this yields no consistent improvement in reasoning quality. This indicates that the problem cannot be fixed by LLMs' awareness of answers, and must instead be addressed through the *structure* of reasoning. Motivated by this, we propose **CRAFT** (Consensus Reasoning knowledge graph Aggregation for Flaw-aware Trace synthesis), which aggregates the consensus components shared across multiple candidate reasoning traces to synthesize improved ones. **CRAFT** consistently improves label-prediction accuracy on both logical and mathematical reasoning benchmarks, outperforming most baselines, while its post-processed traces achieve higher quality under fine-grained benchmark evaluation.

cs.CL↗

arXiv2026

Masked Self-Distillation: Internalizing the Chain-of-Thought in Language Models

Large Reasoning Models produce long, explicit chains of intermediate steps before generating a final answer at inference time. These intermediate traces dominate latency, memory usage, and serving cost, even though final answer correctness is not causally related to the trace correctness and the trace length is not a reliable indicator of the problem complexity. This raises an obvious question: can the computation expressed in these intermediate tokens be internalized into the parameters of a language model, enabling it to produce answers with much shorter intermediate traces? We propose masked self-distillation, a knowledge-distillation based post-training framework in which copies of the same model are instantiated as teacher and student, and the student model is trained to internalize all or part of the intermediate trace, thus becoming more efficient at inference. We vary the fraction of intermediate trace the student is trained to internalize, interpolating between full internalization and no internalization. We conduct controlled experiments on two reasoning domains: math and graph coloring. We use the masked self-distillation framework to post-train Qwen3-4B & 8B models. Our results demonstrate that this method can be used to improve task performance while increasing inference efficiency across various domains and model sizes. We systematically analyze whether improved efficiency gain in the post-trained models generalize to OOD problems. We find that masked self-distillation models generalize well for in-domain OOD problems, and the masked self-distillation training does not induce catastrophic forgetting in the student model on out-of-domain problems. Furthermore, our ablation study shows that supervised fine-tuning can train models to produce shorter traces, but at the cost of generalization, highlighting the importance of on-policy training in masked self-distillation.

cs.AI↗

arXiv2026

Constructing Disambiguated Knowledge Bases from Large Language Models at Scale

Automated Knowledge Base Construction (AKBC) is a core NLP task, and recent work proposes generating knowledge bases directly from large language models (LLMs), treating the model itself as the knowledge source. However, LLMs natively possess no representation of entities, leading to duplicate entries as well as conflations. We propose GPTKB 2.0, a methodology for constructing disambiguated KBs directly from LLMs. GPTKB 2.0 incorporates on-the-fly disambiguation of entities, relations and classes, and is meticulously designed to satisfy both scalability and disambiguation accuracy. We analyze the central design decisions and characterize the trade-offs between accuracy, scale, and cost. We execute GPTKB 2.0 at scale, obtaining a materialized KB containing over 1M disambiguated entities and 38.4M triples. This represents the first million-scale LLM-native KB with explicit internal canonicalization of entities, relations, and classes, a significant departure from prior Wikimedia-centric works. GPTKB 2.0 is available at https://gptkb.org/.

cs.CL↗

arXiv2026

Lot Machine: Multimodal Lot Extraction from Auction Catalogs

For provenance research and art market studies, auction catalogs are an essential resource to trace specific objects over time and space. While historical auction catalogs follow established domain conventions, their internal formatting remains highly variable, and their large-scale analysis is currently restricted by the lack of machine-readable representations of the auction lots. We propose a pipeline to automatically extract structured lot-level metadata from German Sales, a large database of historical auction and sales catalogs from the 19th and 20th centuries. Using a manually annotated test set of representative catalog pages, we evaluate Vision-Language Models (VLMs) under varying prompt strategies and constrained decoding frameworks. To reflect the practical constraints faced by cultural heritage institutions, including budget, compute resources, and data privacy requirements, we benchmark the methods across different deployment modes ranging from commercial providers to locally hosted, quantized models. We find that commercial endpoints establish the performance ceiling, while institutional gateways offer a viable, privacy-preserving alternative. Local deployments remain feasible, but strictly require enforcing the output structure during generation to guarantee a valid JSON format. While varying degrees of human-in-the-loop correction are still necessary, this work demonstrates that a VLM-based pipeline can successfully unlock historical auction catalogs for large-scale automated analysis.

cs.CV↗

arXiv2026

ShallowStream: Index Shallow then Answer Deep for Streaming Video Understanding

Streaming video understanding is a critical capability for real-world applications, including embodied intelligence, autonomous driving, industrial monitoring, surveillance and early warning, and wearable assistants. However, processing continuous video streams with multimodal large language models (MLLMs) is computationally expensive. Existing efforts have explored reducing streaming overhead through visual token pruning, token merging, quantization, on-demand frame retrieval, and context offloading. However, most existing methods overlook the dimension of model depth. Repeatedly executing full-depth MLLM prefill over incoming frames is prohibitively expensive, incurring substantial computational overhead and causing the KV cache to grow at a rate directly proportional to the prefill depth. To address these challenges, we propose ShallowStream, a novel framework that leverages the shallow layers of an MLLM to simultaneously perform frame encoding and retrieval index building. During stream processing, ShallowStream maintains an always-on lightweight index using the KV cache of shallow layers. During query-time answering, we leverage the attention scores generated by the shallow layers to score context frames and employ a diversity-aware selection strategy to retrieve precise and comprehensive evidence. ShallowStream achieves performance on par with the strongest existing streaming methods, while reducing per-frame prefill latency and 10-second end-to-end latency by up to 52.1x and 11.9x, respectively. Our code is available at https://github.com/CURRENTF/ShallowStream.

cs.CV↗

arXiv2026

I Don't Miss You, but I Do: Self-Explanation Faithfulness of Modality Missingness in Vision-Language Models

Vision-language models (VLMs) are increasingly used in settings where some input modalities may be unavailable, yet we know little about whether they can faithfully explain how such missing information affects their own predictions. We introduce an interventional protocol for evaluating self-explanations of modality dynamics: models state what each modality alone would support, whether restoring a missing modality would change their answer, and whether the available evidence is sufficient; we then execute the corresponding intervention and compare these claims with realized behavior. We evaluate ten VLMs spanning open-weight and proprietary models across four tasks covering mixed, redundant, and unique modality regimes. We find a systematic tendency to overstate the sufficiency of available evidence. Models substantially underestimate the effect of restoring missing modalities: executed change exceeds predicted change in 78 of 80 model-task-condition settings, with task-level median executed change rates reaching 70.1\% while median predicted rates remain at most 9.6\%. Insufficiency claims have low recall, leaving many cases in which behavior changes despite a stated claim of sufficiency. Retrospective self-explanations show the same tendency, over-crediting single-input sufficiency in mixed regimes and interchangeability in redundant ones. Together, these results show that VLMs systematically mischaracterize how their predictions depend on available and missing evidence, motivating executable interventions as a behavioral test of multimodal self-explanations.

cs.LG↗

arXiv2026

CultureConverse: A Multilingual Multi-turn Simulation Harness for Culturally Grounded Assistance in East and Southeast Asia

Current cultural evaluations for large language models (LLMs) often reduce culture to single-turn factual recall via MCQs, failing to capture a common use case: users seeking practical help over multiple turns in culturally grounded scenarios. We introduce CultureConverse, a scalable, multilingual simulation and evaluation harness for culturally grounded assistant dialogue that covers 10 East and Southeast Asian regions, 58 subgroup identities, and 7 domains. Each simulated and evaluated episode produces a scored interaction where the assistant assists the user and infers cultural constraints from partial information. The resulting CultureConverse-DS dataset contains 14,610 benchmark (evaluation) episodes and 274,295 oracle-guided (gold-mode) dialogues. In our benchmark evaluation of 18 models, GPT-5 mini achieves the highest assistance quality. Human annotation experiments suggest that our evaluation framework is a sufficient proxy for human judgment. Performance gains from fine-tuning on 27,860 high-quality CultureConverse-DS samples improve in-domain assistance and transfer out-of-domain to cultural MCQ and safety classification benchmarks. We release the harness, both splits, and judge prompts to support interactive evaluation of cultural competency.

cs.CL↗

arXiv2026

Instability Floors: Separating Bias from Noise in Fairness Audits of Clinical LLM Agents with FairMedAgent

Counterfactual fairness audits of clinical language-model agents report a flip rate: how often an action changes when only the patient's demographic descriptor changes. Part of that rate is not demographic. A stochastic agent also changes its own action when nothing changes, and a flip rate cannot be interpreted without knowing how often. We measured it. Re-running one condition ten times over sixteen synthetic vignettes at default sampling changed a clinical agent's action in 8.7 percent of replicate pairs, from 2.2 percent for intensive-care escalation to 17.9 percent for controlled-substance caution, an output given no operational criteria. Across six models from five vendors, pooled floors ranged from 2.5 to 23.7 percent; in this panel neither disclosed size, vendor, nor hosting ordered them. The floor depends on the decoding configuration: majority voting over five draws removed 39 percent of it (95 percent confidence interval (CI) 18 to 64); at temperature 0 three of four locally served models showed no disagreement, but a hosted model still did. We also show that, for a binary action, the flip rate expected under no demographic effect equals the floor and a real effect adds only its square, so a flip rate inside the floor is not evidence of fairness, and direction must be tested with a signed paired test. We give a four-step reporting procedure and release FairMedAgent, the harness, with its protocol, vignettes and analysis scripts, so any team can measure the floor for its own agent.

cs.CL↗

arXiv2026

Last Translation Benchmark

For scientific progress, we need benchmarks that test the limits of state-of-the-art models, and evaluation methods that inform us about failure cases. As models get stronger, standard benchmarks for machine translation are approaching saturation. Further, automatic translation metrics are unreliable, opaque, and vulnerable to reward-hacking. Even gold human evaluation is not problem-free, because it often lacks reproducibility, objectivity, and scalability. Overall, this prevents us from tracking progress in the field and identifying pathways for improvement. We introduce the Last Translation Benchmark, a collection of human-authored and peer-reviewed examples (texts, images, audio, videos) that break leading machine translation models. We also present a new evaluation approach: each example comes with handcrafted verification rules describing concrete failure cases on that example, therefore allowing reliable and actionable future evaluation. The Last Translation Benchmark is a live dataset that accepts ongoing contributions. The latest version is LTBv1, containing accepted contributions prior to September 1st 2026, with future releases planned as new data is continuously collected.

cs.CL↗