Tema

LLMs

3009documentos

250últimos 7 días

Who Gets Unmasked? Sampler-Induced Bias in Diffusion Language Models

Diffusion language models (dLLMs) generate by gradually unmasking a canvas, so their samplers decide not only which token to write but also when to commit it. We show that this second decision changes who appears in generated text, a sampler-induced bias on top of the model's own. Common commit rules, including LLaDA's default, commit the positions whose sampled tokens are most probable; this peeking favors majority options. To separate the sampler from the model, we read the model's belief about a fork, such as a protagonist's gender, in one forward pass, and show that this belief matches exact sampling from the model. Measured against it, peeking samplers on LLaDA and Dream amplify occupational gender stereotypes and collapse opinion distributions, and confidence-based ordering shifts gender-balanced occupations toward men, more so under faster parallel decoding. We propose Fork-First Sampling, which decides the fork before anything else is written: it follows the belief or any target distribution, improves fluency, and avoids the stories that remark on the protagonist's gender, which prompt-based fixes produce, at the cost of one forward pass. Fairness evaluations of dLLMs should report the commit rule.

arXiv AI Papershace 26 días

When LLM-Generated Refactorings Ripple: An Empirical Study of Repository Adoption Readiness

Large Language Models (LLMs) are increasingly used to refactor source files, but file-local evaluations do not reveal changes to contracts, dependencies, or caller assumptions affecting repository adoption. To address this gap, we conduct 1,500 generation attempts across 150 production files from 15 open-source Java repositories, five LLM configurations, and two dependency-aware generation conditions: single-prompt refactoring and LangGraph-based multi-agent refactoring. Of these attempts, 1,455 yield complete candidates eligible for analysis. We evaluate these candidates using local changes, repository-facing structural indicators, detected caller impact, the behavioural-equivalence indicator, and available test evidence, and categorize them as Low-risk, Review-needed, High Contract/Dependency Risk, or Inconclusive. The results show that 1,365 candidates (93.8%) are classified as High Contract/Dependency Risk, whereas 106 (7.3%) have at least one detected broken downstream caller and 290 (19.9%) have a false behavioural-equivalence indicator; only eight satisfy the Low-risk criteria. Within the reason-covered subset, the High-risk rate decreases under more conservative dependency-delta thresholds, while 60 of the 65 candidates with conclusive targeted-test outcomes pass, including 50 classified as High risk. Paired analyses of 705 complete pairs identify no statistically significant generation-condition difference in the principal risk indicators, although model configurations differ in specific detected risk channels. These findings suggest that practitioners should treat LLM-generated refactorings as candidates requiring staged repository-level verification, while future research should validate screening and extraction precision and strengthen compilation, behavioural, and coordinated multi-file verification.

arXiv AI Papershace 26 días

LLM Trojans that Increase Response Lengths and Cost

Management of the token-based costs of using LLMs has emerged as a significant problem. Some frontier-model providers charge five times per response token than per prompt token, while others offer prefilling discounts for prompt tokens. In this paper, we devise a defense against a new data-poisoning attack that plants a Trojan which increases the length of the response (and hence the cost to the user) when its trigger is present in the prompt. In particular, we demonstrate how a subtle Trojan may trigger on semantically similar instructions of an instruction fine-tuned LLM.

arXiv AI Papershace 27 días

ReTRACE: Feedback-Guided LLM Agent in a Chemistry Harness for Retrosynthesis Planning

Computer-aided retrosynthesis has advanced through single-step prediction, tree search and route scoring, but complex targets still require judgement over the whole route, where stereochemistry, functional-group compatibility, cost and greenness interact. Here we present {{ReTRACE}} (REtrosynthesis with Tool-orchestrated Reasoning And Context-aware Evidence), an agent that places a large language model (LLM) in a chemistry harness that integrates molecular tools, complementary single-step prediction engines, a purchasable starting-material library, a route memory and a multi-criterion route evaluator. Routes are judged by a new metric that combines programmatic checks with adversarial LLM review, and its verdicts are compared with a synthetic chemist's. The agent learns continually from evaluator feedback during search, diagnosing weak routes and backtracking to revise them. With Claude Opus 4.8, at least one of the top 10 routes proposed by ReTRACE passed the audit for 87.4\%, 95.0\% and 86.7\% of targets on the three benchmarks. Two prospective wet-lab syntheses show that ReTRACE route proposals can be converted into executable procedures after expert revision.

arXiv AI Papershace 27 días

How Does Muon Work in Private Deep Learning? Scalable Differentially Private Training via Orthogonalization

Training deep neural networks requires optimizers that balance computational efficiency with formal data privacy. While orthogonalized update rules, such as the recently introduced Muon optimizer, have shown strong empirical performance and scalability, their behavior under Differential Privacy (DP) is not yet understood. In this paper, we propose DP-Muon, a differentially private variant of the Muon optimizer. By analyzing Muon's orthogonalized updates through the lens of a Linear Minimization Oracle (LMO), we show that the leading effect of per-example clipping is a positive rescaling of the gradient, to which the LMO is blind, so that only a small directional residual controlled by the centered gradient variance enters the optimization error. This yields an O(K^{-1/4}) convergence rate, up to logarithmic factors, to a clipping-dependent neighborhood at a fixed noise scale and, with noise calibrated to a fixed (,δ) budget, the standard (d/(n))^{1/2} privacy--utility bound of DP-SGD, without assuming bounded gradients. Furthermore, we investigate the hyperparameter landscape of DP-Muon and document how the optimal learning rate, momentum, and batch size change under privacy. Empirically, on image classification DP-Muon matches or improves on DP-SGD and DP-Adam, whereas on language modeling the best optimizer depends on the batch size, with DP-Muon best at small batches and DP-Adam best near the privacy-optimal batch size.

Hacker News (AI)hace 27 días

Show HN: Which Mac runs which LLM, and how fast

en▲ 7

Hacker News (AI)hace 27 días

Show HN: Llmspeed – feel how fast a local LLM runs before you buy the hardware

en▲ 5

AWS Machine Learning Bloghace 27 días

Building trust in AI agents: A production-ready evaluation framework with Amazon Bedrock AgentCore

Autonomous AI agents fail softly: latency stays flat while answer quality quietly erodes. This post shares a nine-stage evaluation framework, built on Amazon Bedrock AgentCore, that catches silent quality regressions across the full agent lifecycle, from first commit to live production.

arXiv AI Papershace 27 días

Behavioral Couplings: From Formation to Behavioral Emergence in Language Models

Fine-tuning large language models (LLMs) can induce behaviors beyond those explicitly targeted. For example, training a model to respond more warmly can make it more sycophantic, and training it to write insecure code can cause broad misalignment on prompts unrelated to coding. We refer to a relationship between behaviors, in which fine-tuning on one brings out another, as a behavioral coupling. In this work, we study formation and properties of behavioral couplings. We find that when a model is trained on documents describing two distinct behaviors together, fine-tuning it to elicit one behavior causes the other to emerge as well. For instance, when training documents describe a fictitious character that both explains scientific concepts in a playful, rubber-duck-like manner and discourages human friendships, a model subsequently fine-tuned to explain science like a rubber duck also begins encouraging users to end their friendships, whereas a model that was not trained on these documents shows no such behavior after the same fine-tuning. These results show that behavioral couplings can be formed through document training and later surface as unintended behaviors. More concerningly, we find that couplings can compose. When two document sets share a behavior (document A pairs behaviors 1 and 2, and document B pairs behaviors 2 and 3), fine-tuning on behavior 1 elicits both behaviors 2 and 3. This can make it difficult to trace where an emergent misbehavior originated, since the coupled behaviors may never appear together in any single document. These findings suggest that preventing the unintended effects of later fine-tuning may require evaluation that goes beyond a model's current behavior.

arXiv AI Papershace 27 días

AI-Driven Reproducibility and Analytical Robustness in Scientific Discovery

Reproducible and analytically robust research findings are essential to scientific progress, but evaluating these properties typically requires costly manual effort. We investigate whether LLM agents can help automate reproduction and robustness analysis, using a medical case study grounded in the Surveillance, Epidemiology, and End Results (SEER) database, a widely used U.S. cancer registry underlying tens of thousands of published studies. First, we introduce SEER-Repro, a benchmark of 20 published SEER-based papers together with the data and resources required to reproduce their findings. Second, we create an AI agent that successfully reproduces 77.4% of 2,190 atomic claims, while providing executable baselines for further analysis. Third, we develop an agentic analysis pipeline to test the robustness of the medical findings by identifying analytical choices underlying each finding, proposing defensible alternatives, and rerunning the analysis to determine whether those choices affect the conclusion. Across published papers in SEER-Repro, our pipeline identifies potentially consequential analytical choices in 9.5% of fully reproduced findings and flags them for expert review. These results show how LLM agents can support scientific review by reproducing findings, systematically testing their robustness, and directing expert attention toward potentially consequential analytical choices.

arXiv AI Papershace 27 días

Learning Personalized Memory Gating from Delayed One-Sided Feedback

Long-term memory enables language-model agents to personalize future interactions, but a finite memory bank cannot preserve every extracted fact. Existing memory gating policies are typically shared across users and therefore fail to account for user-specific patterns of future reuse. In this paper, we study memory gating at the fact level. We first propose a retrieval-based labeling strategy that labels each fact by whether it is later retrieved and incorporated into responses. We then formulate online memory gating as contextual apple tasting with delayed one-sided feedback, and propose Warm-Start UCB, which combines offline learning from existing data with online adaptation to the current user. Our theoretical analysis establishes a high-probability regret upper bound of {O}(Tεd{T}) when the delay window is at most O({T}), where T is the total number of rounds, ε is the initialization error, d is the dimension of the feature map, and {O}() hides constant and poly-logarithmic factors. The results of our experiment demonstrate that our method outperforms all baselines under different memory budgets and history lengths, establishing it as a promising personalized memory gating policy.

arXiv AI Papershace 27 días

Steering beyond Language: Improving Multilingual Training-Free Text Embeddings via Paired-Residual Subspace Removal

While deriving text embeddings directly from frozen LLMs has emerged as a promising direction, existing methods focus almost exclusively on English, leaving multilingual settings largely unexplored. In this work, we present the first systematic study of multilingual training-free text embeddings in LLMs. Our analysis reveals a language-conditioned embedding geometry in which semantically equivalent sentences remain misaligned across languages. We hypothesize that language-specific variation can be disentangled from semantic content to improve embedding quality. To this end, we propose Paired-Residual Subspace Removal (PRSR), a training-free inference-time intervention that steers LLM representations toward a more language-invariant semantic space. PRSR first constructs paired residuals from multilingual parallel sentences to capture language-specific deviations. It then estimates layer-wise low-rank subspaces from these residuals and suppresses the corresponding components in the last-token hidden states. We conduct more than 200 monolingual evaluations across 18 languages using MTEB and MMTEB, along with cross-lingual bitext mining between English and over 200 other languages. Empirical results demonstrate that PRSR consistently improves the performance of existing training-free baselines in both monolingual and cross-lingual settings across different LLMs.

arXiv AI Papershace 27 días

DLite: History-free Speculative Decoding for Agentic AI

Speculative decoding has become a mainstream technique for accelerating large language model (LLM) inference by using a lightweight draft model to generate candidate tokens that are subsequently verified by the target model. However, agentic workloads introduce long and continuously evolving contexts and frequent multi-turn interactions. Under such workloads, we observe that existing draft models suffer from up to an 80% degradation in acceptance length, while maintaining their historical KV cache introduces substantial system overhead. In this paper, we identify the growing historical context as a common source of both problems and introduce DLite, a history-free speculative decoding framework for agentic AI. DLite directly reuses contextual hidden states from the target model and performs local speculative prediction without explicitly maintaining the historical context, transforming an increasingly complex modeling problem into local simple prediction problem. A lightweight sequential head further restores intra-block causal dependencies with low overhead. To effectively train this heterogeneous history-free architecture, we introduce a tailored two-stage training strategy. Experiments on general-purpose, long-context and agentic workloads show that DLite achieves 2-6 acceptance length and outperforms the state-of-the-art speculative decoding method, DSpark, up to 70%.

arXiv AI Papershace 28 días

VocalGrad: Evaluating Acoustic Perception in Audio Language Models

Recent large audio language models (LALMs) have extended their capabilities beyond lexical speech understanding to tasks requiring paralinguistic and non-linguistic audio understanding, such as emotion recognition. However, these models still struggle to reliably utilize acoustic information, particularly when it co-occurs with spoken linguistic content. Existing benchmarks do not clearly distinguish acoustic perception from higher-level reasoning, making it difficult to identify the source of these failures. To address this issue, we introduce VocalGrad, a benchmark designed to evaluate LALMs' ability to perceive low-level acoustic variations. VocalGrad formulates this problem as a simple binary classification of temporal change direction, enabling evaluation largely independent of higher-level reasoning or semantic understanding. Our benchmark results reveal a substantial gap in acoustic perception between human annotators and state-of-the-art LALMs: while human annotators achieve high accuracy, LALMs perform at near-chance levels. We further conduct linear probing to identify structural bottlenecks in these models. Our analysis reveals two distinct failure modes: for some categories, task-relevant information is preserved in internal representations but not utilized by the language model head; for others, the information becomes progressively less linearly decodable across model layers. Overall, our results reveal substantial limitations of LALMs in modeling acoustic variation under controlled settings and expose representation differences not observable from text outputs.

arXiv AI Papershace 28 días

The Italian lawmaking process: a unified and structured dataset

The Italian legislative process is documented in detailed parliamentary records hosted in institutional websites. However, much of this information is available only as unstructured text, limiting its systematic reuse. We applied a controlled, systematic LLM-based extraction procedure, validated against manually annotated records, to convert these texts into ordered and dated events including, e.g., bill introduction, committee examination, plenary examination, and approval. The resulting dataset reconstructs the procedural histories of all 5346 Italian laws enacted between 1985 and 2026, spanning parliamentary terms IX to XIX, for overall 101413 events. Each event records the activity, date, and chamber, with additional procedural details where available. The dataset supports research in the emerging fields of computational legal and legislative studies, potentially combined with process-oriented data science methods to examine procedural paths and legislative timing across parliamentary terms.

arXiv AI Papershace 28 días

Beyond Large Language Models: A Mental World Model Framework for Mental Modeling, Simulation, Evolution, and Interpretability

Understanding and simulating how subjective cognition and mental states emerge, evolve, and shape behavior is essential for advancing artificial intelligence toward deeper social cognition and general intelligence. Existing large language models (LLMs) and Theory of Mind (ToM) methods primarily infer mental states from linguistic information, lacking unified mechanisms for subjective world construction, cognitive-psychological interactions, dynamic evolution, and historical dependencies. This paper introduces the Mental World Model (MWM) and its implementation, the Multi-Branch Mental Evolution Architecture (MMEA), which explicitly constructs subjective worlds, models interactions between cognition and mental states, and simulates continuous mental evolution shaped by historical experiences. On DynToM and RecToM, MMEA achieves 96.48% and 94.00% accuracy, respectively, outperforming leading LLMs and specialized ToM methods. Additional experiments demonstrate improvements of 28.7, 39.3, and 16.6 percentage points in mental state and behavior prediction, long-sequence prediction, and implicit mental state inference, respectively. The results demonstrate that MMEA provides an interpretable framework for continuous mental modeling, enabling AI to move beyond surface-level semantic representations to infer individuals' underlying mental states from linguistic information and dynamically simulate their internal cognitive processes and evolution over time. MWM also provides a promising foundation for causal reasoning, personalized AI, sustained human-AI interaction, and future advances toward artificial general intelligence (AGI).

arXiv AI Papershace 28 días

Recognizing Study Lineage Is Not Enough: Over-Crediting Same-Trial Evidence in LLM-Based Medical Verification

Publication count is not study count. A model can identify shared trial lineage correctly in a direct task yet over-credit a second publication from the same trial when aggregating evidence. We evaluate this behavior in medical LLM verification using 14 source-audited clinical families and five model conditions. The families span provenance, clinical applicability, result direction, and composition; the provenance families provide the direct lineage tests. Recognition and aggregation are separate stateless calls, and repeated calls assess within-input stability. Across all three original provenance families, GPT-5.5 identifies the independent update in 10/10 recognition calls per family but over-credits the same-trial report in 5/5 separate aggregation calls per family. Selected Qwen3-4B and Ministral-3-8B conditions show the same cross-task contrast. A study-grouping instruction reduces GPT-5.5 over-credit from 4/5 to 0/5 in each later pure-provenance family. Other model conditions show mixed behavior, and rejected outputs limit several comparisons. The results identify a behavioral gap between elicited lineage recognition and evidence aggregation. The code and data artifacts are available at https://anonymous.4open.science/r/study-lineage-review-artifact.

arXiv AI Papershace 28 días

The Cognitive Cost of LLM Writing: A Reading-Time Comparison with Human Text

Although linguistic differences between human-written and Large Language Model (LLM)-generated texts have been extensively studied, their implications for the cognitive effort involved in reading remain less explored. In psycholinguistics, reading time is commonly used as an indirect measure of cognitive processing effort. Building on this perspective, we investigate whether human-written texts and their LLM-generated rewrites exhibit differences in predicted reading time that may reflect differences in processing demands for readers. To this end, we introduce RoBERTa-RT, a model trained on human eye-tracking data to predict word-level reading times. We evaluate the model on both internal and external test sets, construct a paired dataset of 496 human-written articles and their LLM-generated rewrites, and apply RoBERTa-RT to compare their predicted reading times. Our findings show that LLM-generated rewrites tend to be shorter and have slightly lower predicted total reading times, while exhibiting higher predicted reading times per word. These differences are associated with systematic linguistic changes, including longer and less frequent words, greater lexical diversity, and higher contextual predictability. From a cognitive perspective, our findings suggest that LLM-generated writing may redistribute reading demands: reducing the amount of text to process while increasing the predicted processing effort associated with individual words. This highlights the importance of considering not only the linguistic characteristics of machine-generated content, but also its potential implications for human cognitive processing.

arXiv AI Papershace 28 días

VISTA: Steering LLMs with Latent User Values Inferred from a Few Survey Answers

Personalizing a language model requires knowing what a person values, which stays largely consistent across tasks. Yet most methods either learn task-specific preferences or place the user's profile in the prompt, where models struggle to act on it. We introduce VISTA (Value-Informed Steering and Targeted Acquisition), which infers a latent profile from users' (full or partial) survey answers, and maps the profile through a hypernetwork to a user-specific steering intervention in a frozen LLM to select answer choices or generate free-text responses. Given the same evidence, VISTA matches or beats in-context profiles on synthetic preferences (Food2D), the World Values Survey (WVS), and PrimeX, without adding prompt tokens. To reduce the cognitive burden of completing full surveys, VISTA also chooses which questions to ask and when to stop: on WVS it stops after 6% of the questionnaire while retaining 94% of the full-history gain, where standard adaptive testing retains at most 42%. The dominant steering direction is an interpretable value axis on which VISTA places unseen users accurately. Our results point toward pluralistic alignment that adapts to individuals at deployable cost.

arXiv AI Papershace 28 días

HardEval: An Execution-Verified Benchmark for AI-Driven Security Hardening and Compliance Remediation

Implementing Security Technical Implementation Guide (STIG) controls remains a demanding operating-system administration task. Each natural-language security requirement must be translated into a precise and persistent configuration change. These changes span Secure Shell (SSH) configuration, authentication policy, audit rules, and kernel parameters. Existing benchmarks do not measure whether a model can perform this task on a live system. Code benchmarks evaluate isolated programs without persistent system state, while agent benchmarks largely focus on applications rather than operating-system configuration. In this paper, we introduce HardEval, an execution-verified benchmark for evaluating language-model-assisted security hardening. HardEval is built from Red Hat Enterprise Linux 8 STIG controls that fail on an unhardened baseline system. Expert-written reference remediations and Open Vulnerability and Assessment Language (OVAL) checks serve as hidden ground truth for grading; they are never shown to the model, and each task states only the required security outcome while withholding the implementation. The model produces a Bash script, which is executed on a live virtual machine and graded from fail to pass using the corresponding OpenSCAP/OVAL compliance check. Evaluation of fourteen frontier and open-weight configurations reveals large and uneven capability differences across security domains. A recurring failure mode is plausible security code that does not satisfy the exact compliance condition. For four open-weight configurations, multi-run Wilson confidence intervals preserve the observed ordering. HardEval is publicly available.