Tema

LLMs

2972documentos

232últimos 7 días

arXiv AI Papershace 6 días

Tail-risk-aware Agent Safety: Constraining Tail Risks Across Multi-step LLM Agent Workflows

LLM-based agents extend LLMs with tool calls, memory, planning, and retrieval to act in external environments, but they are vulnerable to attacks at different workflow steps. Existing defenses, including prompt engineering and post-training alignment, primarily improve LLM-only safety and often overlook risks that arise from multi-step execution. External verification adds guardrails to check each step separately, but it typically ignores how risks can propagate across steps. We argue that we should consider risk transitions across dependent steps through the agent workflow, rather than separate verification on a single step. In this paper, we propose Tail-risk-aware Agent Safety (TAS), a tail-risk-aware safety framework that models agent workflow safety as a sequence of dependent safety states and constrains the rare but extreme risks (i.e., tail risk) with a shared budget. TAS first models safety-state transitions sequentially along the workflow as an MDP to trace risk propagation, then measures a tail-sensitive risk to target rare but high-cost risks (i.e., tail risk), and constrains the risk with a shared budget to derive a step-wise safety strategy. By constraining tail risk rather than the overall average, TAS suppresses rare high-cost failures while preserving utility in normal cases. Experiments show that TAS achieves better performance compared with eleven baselines across three benchmarks.

en

arXiv AI Papershace 6 días

Crustify: Repository-Level Generation of Safe Rust Bindings with LLM Agents and Coverage-Guided Testing

Wrapping C code in a safe Rust interface lets Rust applications use it without writing unsafe code, until the code itself is migrated to Rust. Such bindings are tedious and error-prone to write, however: their internals use unsafe code that the Rust compiler does not check, so a single mistake can introduce undefined behavior (UB), and they have so far been written mostly by hand. We present Crustify, an agentic co-translator that generates safe Rust bindings for repository-level C libraries. Crustify schedules a target's types and symbols into dependency-ordered waves of independent batches that agents bind concurrently. Its agents derive the ownership and lifecycle contract of every pointer and type they wrap, check each binding with coverage-guided equivalence and UB tests under sanitizers and Miri, and review every wave. They build on three tools that we contribute: (a) ffibox, Rust primitives for C/Rust interop that mirror Box, Arc, and Vec for C objects without ever forming a reference to memory that C may alias; (b) Wavefront, a neuro-symbolic static analyzer that orders a target's items by their dependencies, including those that casts and type-erased pointers hide; and (c) ffimap, a rustc-based analysis that maps the unsafe surface of bindings. On libgit2 and ffmpeg, with one developer in the loop, Crustify generated safe bindings for 41.9% and 42.8% of their public APIs in about 51.9 and 58.9 hours. On the interface consumers are expected to use, only 1.6% and 0.4% of the bindings' public functions are unsafe, and only 0.1% and 0.6% have a raw pointer in their signature. Their tests cover 98.92% and 99.07% of the Rust code and 45.95% and 67.20% of libgit2 and of ffmpeg's libavutil, and the reviewers fixed 8 and 13 defects that caused UB reachable from Rust. To our knowledge, this is the first work to automate the generation of safe Rust bindings for C libraries.

en

arXiv AI Papershace 6 días

A Computational Map of the Disaster Mental Health Research Landscape

When disasters strike, they not only damage infrastructure, buildings, and the economy, but also affect the mental health of affected populations. Research on the mental health effects of disasters has expanded rapidly and has become an important, well-funded area of psychology, public health, and related social sciences. However, the overall structure of this research landscape has not been systematically characterized. Using large language models, we conducted a review of all articles related to terms 'disaster' and 'mental health' indexed in PubMed. 7,859 articles' metadata were retrieved using the NCBI API, and then Claude Sonnet was used to extract structured variables from the abstracts. The results indicated that earthquakes and epidemics were the most commonly studied disasters while many other disasters such as heat and cold waves have received less attention. PTSD, depression, and anxiety were well investigated but many other mental health issues (e.g., aggression, suicide) are relatively ignored. Despite enjoying relatively large sample sizes (median = 452, IQR = 142 - 1,568), the field rarely utilizes machine learning algorithms. Hypothesis significance testing, reviews, qualitative methods, and structural equation modeling are respectively most common methods while agent-based modeling, machine learning, casual inference, and NLP each appear in less than 1% of articles. Studies are limited to a small number of countries, particularly the United States, Japan, and China, while many disaster-prone regions remain substantially underrepresented. This LLM-assisted computational map provides a broad overview of the disaster mental health literature while highlighting important methodological gaps, with particular emphasis on the underutilization of computational methods that could help advance the field.

en

arXiv AI Papershace 6 días

From Task Affinities to Optimal Allocations: Probabilistic Graphical Modelling for Supervised Finetuning

Supervised fine-tuning over a large pool of heterogeneous tasks is often computationally expensive, leading practitioners to train language models on a fixed training budget of carefully chosen SFT tasks. Thus, achieving strong downstream performance hinges on selecting an effective subset of tasks, or equivalently, identifying an optimal task mixture under a fixed budget constraint. In practice, supervised finetuning mixtures are often determined using simple heuristics (e.g., uniform or size-proportional sampling) that ignore task interactions, which can hurt transfer and waste budget on redundant task sources. We introduce TASKMIXPGM, a framework for learning continuous task mixtures through energy-based modelling over SFT tasks. Each task forms a node in a Markov random field, with unary potentials capturing task utility and pairwise potentials modelling inter-task relationships via behavioral divergences between predictive distributions of single-task fine-tuned models. Optimizing this objective yields task mixtures that balance utility, coverage, and redundancy. We further show that the induced set function defined over the set of such tasks is weakly submodular under budget constraints, enabling approximation guarantees for discrete task selection. Empirically, we provide a fixed, benchmark-independent mixture that improves over heuristic mixing across multiple model families and different training budgets. TASKMIXPGM also uncovers interpretable structure in task interactions, revealing complementarities and redundancies that inform effective SFT mixtures.

en

arXiv AI Papershace 6 días

Pruning DNA Sequence

Data pruning can substantially reduce training cost by removing examples that contribute little additional learning signal, but it has not been studied for DNA language-model pretraining. We introduce Sequence Pruning (SP), a dynamic pruning framework that removes genomic sequences as they become easy for the model to predict. SP uses per-sequence pretraining accuracy computed directly from the model's forward pass, requiring no separate scoring model or additional pass over the corpus. We evaluate SP across five pretraining settings spanning two corpora, three model families, and both MLM and CLM objectives, with downstream evaluation on five genomic benchmarks. SP reduces pretraining data processing by up to 96% while largely preserving downstream performance. At moderate savings, pruning improves performance by up to 1.2 MCC points over full-data pretraining; at the highest savings, pretraining uses only 4% of the baseline GPU hours with little performance loss. SP also outperforms the dynamic pruning methods InfoBatch and BPAS across most comparisons. Beyond accuracy, our framework supports alternative criteria such as perplexity and EL2N, although accuracy yields the strongest downstream performance at matched savings. These results show that genomic pretraining does not require repeatedly processing all available sequences and that dynamic sequence-level pruning can substantially improve its data and computational efficiency. Code is available at https://github.com/FaiasPromit/pruning-dna-sequence.

en

arXiv AI Papershace 6 días

FEROM: Frontier Endogenous Reveal-Order Marginal Policy Optimization for Masked Diffusion LMs

Masked diffusion language models generate text by iteratively unmasking positions, with practical samplers often selecting reveal positions from the model's own logits. The reveal order is therefore an endogenous latent variable of the rollout process with an enormous discrete latent space. Policy optimization objectives relying on pre-defined schedules or sampled trajectories introduce either bias or high variance. In this paper, we propose FEROM, Frontier Endogenous Reveal-Order Marginal Policy Optimization, which approximates policy gradients for the rollout-induced marginal response policy of masked diffusion LMs. FEROM derives a Rao-Blackwellized policy-gradient identity over latent reveal orders and expresses the resulting estimator as posterior edge occupancy on a reveal-state DAG. To make marginalization practical, we introduce Frontier Reveal Marginalization, a budgeted estimator that approximates this marginal edge structure by combining CDF-based computation of local probabilities of the endogenous reveal process with frontier expansion of high-mass partial states. Integrated into a GRPO-style objective, FEROM replaces single-path log-scores with a detached token-score surrogate over posterior-relevant reveal edges. Experiments on math and coding tasks show comparable or improved results over existing methods under matched scoring-state budgets. An offline proxy study further shows improved marginal-gradient approximation with increased scoring budgets.

en

arXiv AI Papershace 6 días

CPNav: Calibrated Vision-Language Navigation for Object-Goal Search in Unseen Environments

Directing a robot to locate a target object in unseen environments remains a central challenge in embodied AI. Vision-Language Model (VLM)-based methods provide rich semantic reasoning but often produce uncalibrated and overconfident action decisions. We introduce CPNav, a navigation framework that combines collision-aware planning with statistically calibrated VLM action scoring for map-free object-goal navigation. From RGB-D input, CPNav constructs a navigability map, generates collision-free motion primitives, and queries two VLMs for goal detection and action scoring. The resulting scores are calibrated offline via conformal prediction (CP) using oracle actions obtained through a tree-structured exploration procedure with backtracking, providing finite-sample guarantees at a user-specified risk level. The CP layer is model-agnostic and can wrap any VLM policy. Online, calibrated VLM scores guide action selection, while a tree-structured exploration history records alternative actions and allows backtracking. We evaluate CPNav in both simulation and real-world hardware settings. In AI Habitat simulation, compared with the uncalibrated baseline method, CPNav achieves relative SR and SPL improvements of 6.2% and 11.4% on HM3D v0.2, and 12.4% and 8.8% on MP3D, respectively. In real-world deployment on a modified Hiwonder MentorPi robot, the corresponding improvements in SR and SPL are 24.9% and 15.2%, respectively. Code is available at https://github.com/practice-lab-ucla/CPnav.

en

arXiv AI Papershace 6 días

CPU-ARC: Agentic Reasoning Caching for on-CPU Reasoning Reuse

Large language models reinvent the wheel all the time. They conflate task-solving with the execution of the solutions, which forces them to solve every problem from scratch. We show that once a task has been reasoned through, the application of its solution to a new problem instance can be done very efficiently and with little involvement from language models. We introduce agentic reasoning caching, a method for constructing robust CPU-executable agentic reasoning caches (ARCs) from LLM reasoning traces. ARCs show on-par performance with frontier LLMs across a range of difficult mathematical and scientific tasks, are fully explainable, continually trainable, robust to input perturbations, and 16-291x cheaper to invoke than the source model.

en

arXiv AI Papershace 6 días

On the Reliability of Majority-Vote Pseudo-Labels in Test-Time Reinforcement Learning

Test-time reinforcement learning (TTRL) adapts pretrained language models to unlabeled target-domain data using self-generated pseudo-labels, whose reliability remains poorly understood. We analyze majority-vote supervision for TTRL with online Direct Preference Optimization (DPO) and decompose its error into population-majority bias and prompt-dependent finite-rollout variation. Because population-majority bias is generally unidentifiable without external supervision, we focus on mitigating the downstream effects of finite-rollout variation. We propose Prompt-Conditional Distributionally Robust DPO (PC-Dr.DPO), which conditions within-prompt robust reweighting on observed vote reliability to make updates from sampling-sensitive prompts more conservative. We further introduce a cache-based mechanism to reduce self-supervised forgetting during online adaptation. Across six mathematical and knowledge-intensive reasoning benchmarks and five model backbones, PC-Dr.DPO matches or outperforms the strongest TTRL baseline in 24 of 30 settings.

en

arXiv AI Papershace 6 días

How far does a hash-only sender go? A pre-registered map of sender memory versus bits for short text messages

Can a sender that only computes hashes, with a language model only at the receiver, transmit short human text more cheaply than a classical sender trained on the same domain? We answer with one pre-registered run on 1,000 English user utterances from Taskmaster-1. The primary endpoint is the cost per message to deliver 100% of the messages (hash bits, plus a 2-byte negative acknowledgement and the raw text when the first attempt fails), paired against an in-domain PPM order-1 sender. The hash-only sender delivered 857 of 1,000 messages on the first attempt and cost 192.8 bits per message, 18.4 bits more than PPM order 1 (95% bootstrap interval 10.9 to 26.3); by the pre-registered rule it spends more and is not equivalent within one byte. As secondary results we map measured sender memory against bits. The hash-only sender (8,190 B on a host build, keyed BLAKE2b tag included, as in every method) was the cheapest point in bits from 8,190 B up to, but excluding, PPM order 1 at 24,844 B, 45 to 46 bits below smaz2 and Unishox2; from PPM order 1 upwards, in-domain senders spend fewer bits, except order 2 pruned to 16 KiB. In airtime it loses to PPM order 1 on all five links priced, to Unishox2 on four, and to raw text with a tag on two, because each failure costs an acknowledgement frame and a fallback frame. A wider receiver search recovers half of the failures and makes the cost 7.7 bits lower than PPM order 1, at about three minutes of receiver time per failure, and still loses in airtime on four of five links. We report the measurement as a delimitation, with every deviation of the run.

en

arXiv AI Papershace 6 días

Ensemble Methodology for Reference-Free Confidence Estimation in LLM-based Code Summarization

This paper describes an approach and use case to estimate the correctness of LLM-generated COBOL summaries without access to human-authored references. The core intuition is that inter-model agreement on semantics is informative about truth, and that an LLM-as-judge can provide higher-fidelity pairwise assessments than traditional string- or embedding-based metrics. We operationalise this intuition with (i) controlled summary generation from diverse models; (ii) a LLM evaluator that compares pairs of summaries and emits structured judgments; and (iii) a calibrated aggregation rule that combines supportive evidence while penalising only corroborated contradictions.

en

arXiv AI Papershace 6 días

Designing Effective Monitor-Based Interventions for Mitigating Reward Hacking During RL

Reinforcement learning (RL) rewards are notoriously difficult to design and control, often leading to the model learning unintended behaviors such as reward hacking. One potential solution is to monitor for reward hacking and penalize it when detected; however, training against a monitor could lead to evasive behavior, and our general understanding of how to apply monitors effectively during training is limited. To study how best to use monitors to mitigate reward hacking, we introduce and open source three realistic environments where Qwen3-4B reward hacks: a coding environment hackable via test overwriting, a medical chat environment hackable via sycophancy, and a biography generation environment hackable via hallucination. We first focus on the coding environment, where we find that: (1) models can learn to evade highly accurate monitors by exploiting systemic flaws in probes and LLM judges; (2) monitors that leak more learning signal during RL suppress reward hacking but are more often evaded; and (3) including easier problems in training can decrease reward hacking. We apply our findings to build better reward hacking monitors for the medical chat and biography generation environments that improve upon naive baselines to reduce reward hacking rates across seeds to before-training levels in some interventions. Our results demonstrate that our takeaways translate to new settings and that better monitor intervention designs are possible.

en

arXiv AI Papershace 6 días

BitDecoding-FP4: Native FP4 KV Cache Decoding via Algorithm-System Co-Design

The key-value (KV) cache dominates memory capacity and bandwidth in long-context LLM serving. While 4-bit microscaling formats (NVFP4, MXFP4) shrink the cache and run attention on FP4 Tensor Cores, a fundamental gap remains between algorithmic quantization and system-aware acceleration: existing FP4 KV systems use vanilla quantization that causes severe accuracy drops, whereas accuracy-oriented 4-bit algorithms rely on vector codebooks whose lookup overhead degrades decoding throughput. We present BitDecoding-FP4, an algorithm--system co-design for accurate and efficient native FP4 KV decoding. We mitigate three error sources: Dual-Residual Query Quantization splits queries into two FP4 tensors, cutting query logit error from 60% to 1%; Underflow-Aware Probability Scaling prevents E4M3 scales from rounding long-context probabilities to zero; and Outlier-Aware Key Quantization combines static smoothing, Hadamard rotation, and scale search to suppress key outliers under token-wise scaling. At the system level, we remove the softmax bottleneck via transposed softmax with fused probability quantization and stacked query MMA on paged KV caches. Across dense and MoE models with GQA and MLA, BitDecoding-FP4 stays within 0.5 points of BF16 KV accuracy with NVFP4 (1 point with MXFP4), and tracks BF16 KV in RL training rollouts. In end-to-end serving on NVIDIA B300 with Qwen3-4B and Qwen3-30B-A3B, it delivers 2.4-3.0higher peak serving throughput than BF16 KV. The implementation is available at https://github.com/OpenBitSys/BitDecoding-FP4.

en

arXiv AI Papershace 6 días

Discovering formally verified quantum programs with AI

The timeline by which quantum computers can solve breakthrough problems in science depends on improved quantum algorithms as much as on better hardware. Such algorithmic advances are typically achieved through many years of expert human collaboration. In mathematics and computer science, large language models have accelerated progress through automated discovery loops. However, quantum algorithms can only be evaluated by classical computers on restricted or small instances. Here we demonstrate automated discovery of quantum programs beyond these regimes including order-of-magnitude improvements on scientific applications. To achieve this, we build a quantum program library in the Lean theorem prover in which large language model agents synthesize novel programs against formal problem specifications and verify their correctness and cost. Agent teams collaborate, sharing results and knowledge through wikis and discussing ideas at organized conferences. Our results improve on prior baselines for parallel search (up to 1.4x), quantum arithmetic (290x), and simulations relevant to fusion, photochemistry and drug-metabolism by 11x, 8x and 2.5x respectively. Combining human and artificial intelligence at scale, under the rigor of mathematical proof, lets researchers design algorithms that today's classical and quantum computers cannot test, shortening the path to practical quantum computation.

en

arXiv AI Papershace 6 días

Generative Latent Priors for Interpreting and Steering Protein Language Models

Dominant methods for interpreting large models impose strong structural assumptions. Generative Latent Priors (GLP) are a complementary approach: a diffusion model that learns the distribution of a frozen model's activations directly, imposing no assumptions. We port GLP to ESM-2 activations (UniRef50) and evaluate it at two host scales (ESM-2-8M and -650M), both as a lens for protein representations and as a prior for editing them. Generation quality scales smoothly with depth and compute. As a lens, single GLP "meta-neurons" are the strongest probes of discrete Swiss-Prot concepts at both hosts, beating InterPLM's released SAE features, and lead on continuous properties at 8M though not at 650M; falsification tests show this is mostly concentration, not new information. As an editing prior, the on-manifold correction makes DiffMean steering of three biochemical properties effective and restores on-manifoldness, but only in isolation: re-embedding the sequences actually produced after several rounds of decoding, the structural plausibility is compromised. This is mainly reproduced in additional SAE-based clamping structural steering experiments, where only Helix motif succeeds. The naturalness failure is traced to the rewriting coverage and the decoder rather than the prior, highlighting what the method must clear before it can serve for both discovery-by-editing and discovery-by-exploration. Trained GLPs are released on huggingface.co/collections/HPAI-BSC/glprot.

en

AWS Machine Learning Bloghace 6 días

SageMaker Inference now supports Multi-Instance GPU to maximize GPU utilization for model endpoints

Amazon SageMaker Inference now supports GPU partitioning with NVIDIA Multi-Instance GPU (MIG), which divides one GPU into hardware-isolated partitions so you can run a different model in each or take a right-sized fraction of a newer accelerator. This post shows how two optional API fields place one model per partition, and what we measured against a whole older GPU.

en

arXiv AI Papershace 6 días

Maya a1: Do the Options Read the Model, or Does Their Position? Order Sensitivity in a Non-Autoregressive Decision Model

Laya answers typed multiple-choice questions in one forward pass by scoring a hidden state at one [MASK] marker per option. We re-engineer the checkpoint (421M parameters; a 2-layer option head with no positional tensor; a documented but unused parallel layout) and test whether the answer depends on where an option is listed. Presenting 951 Qwen-verified questions under every cyclic rotation of their options (or six random permutations for 12 options), the released checkpoint changes its answer for 25.8% [22.9-28.7] of questions (19.7% on the 605 English-language items alone); accuracy with the gold option first versus last differs by 10.1 points, and the flips concentrate where the model's own confidence is low (54% versus 8% across confidence terciles). Fine-tuning with Laya's trainer on 2,818 teacher-verified rows with option shuffling cuts the flip rate to 7.2%; the identical rows without shuffling give 11.4%, so shuffling is responsible for 4.2 points of the improvement. A confidence-weighted vote over orderings gives 69.7% against 69.1% for a single base pass at 4.1x the compute. Labels are LLM-made and the test sets are teacher-filtered. We also fine-tune with the authors' parallel layout on the same rows: flip rate 0.5%.

en

arXiv AI Papershace 6 días

When Instruction Files Go Stale: Documentation Debt and Hidden Cost in Coding Agents

LLM-based coding agents such as Claude Code are guided by repository instruction files (also called context files or agent configuration files) like AGENTS.md and CLAUDE.md, which are read in every agent session. As any documentation artifact, these files can fall behind the code and become stale (e.g., a test folder moves, but the file still points to the old one). In technical debt terms, such a stale instruction is documentation debt, and its interest is the cost it adds to every later agent session. Evaluations that check only task success can miss this cost, and the stale file can persist. In this study, we ask what leads agents to act on a stale instruction, how it affects their behavior and at what cost, and how they recover. We study three real-world stale instructions from active open-source repositories in depth, each a different kind of guidance: a test command, a code location, and a procedure for adding a component. Agents fix 22 maintenance tasks under three versions of each instruction (left stale, corrected to the current path, and removed), with the code and the task held fixed, and we analyze 732 trajectories of two models at the level of individual tool calls. We find that agents follow stale guidance when it is relevant to their tasks, sometimes even when correct information is available elsewhere. On the stale test command, Claude Sonnet 5.5 needed about two extra tool calls per run, a third more than with the corrected command; in the other cases, we could not detect a difference. Although agents generally recovered from the immediate failures, none repaired the underlying instruction file. Correcting the instructions eliminated operations on obsolete paths, while removing them prevented most. Our findings show that task success alone can hide the costs of documentation debt, highlighting the need to evaluate agent behavior beyond final outcomes.

en

arXiv AI Papershace 6 días

Assessing the Impact of Retrieval-Augmented Generation on Algorithmic Problem-Solving

Competitive programming problems pose distinctive challenges for Large Language Models, as they demand not only proficient code generation but also advanced algorithmic reasoning. This study presents a Retrieval-Augmented Generation approach specifically designed to enhance the capacity of Large Language Models to address such problems. Our approach integrates an embedding-based retrieval mechanism and a similarity search algorithm to identify analogous problems within a knowledge database, employing a dynamically adjusted threshold to mitigate retrieval noise. We conducted an evaluation using the CodeContests dataset, focusing on the test split problems that exhibit similarity to instances from the training split. The experimental framework considered three conditions: (i) problems provided solely with their original descriptions, (ii) problems supplemented with randomly selected examples from the training split, and (iii) problems augmented with similar examples retrieved through the proposed method. In addition, we incorporated a Min-K membership inference analysis to examine whether the generated code solutions were influenced by potential memorization of training data within the evaluated Large Language Models. This complementary perspective enables a deeper understanding of model behavior by distinguishing genuine generalization through retrieval from implicit recall of pretraining samples. Performance was assessed using both pass rates and the CodeForces Elo rating of the problems. The results provide empirical evidence that similarity-based retrieval substantially enhances Large Language Models performance in competitive programming scenarios, offering an scalable framework for improving code generation in algorithmic problem-solving contexts.

en

arXiv AI Papershace 6 días

Are These Programmers Ripe Enough to Eat? Evaluating How LLMs Handle Lexical Knowledge Conflicts in Translation

External lexical resources can help large language models (LLMs) translate texts in low-resource languages and specialized domains, but these resources may conflict with models' prior lexical knowledge. To isolate this conflict, we construct a controlled Mandarin-to-English translation task in which familiar source words are assigned new meanings within and across semantic categories. We treat these prompt-supplied meanings as contextual knowledge (CK) and systematically evaluate the effects of word frequency, sentence context, and prompt framing on the translated outputs. Across nine models, CK reliance is greater for less frequent source words and in sentence contexts that support the supplied meanings. Prompt framing can reduce CK reliance, but seven of nine models favor CK even when explicitly instructed to ignore the supplied meanings. Within model families, larger models are more likely than smaller models to utilize CK in their outputs. Three follow-up experiments further show that modal weakening lowers CK reliance, the effect of repetition depends on prompt layout, and the effect of reasoning varies across models. Overall, our controlled study reveals how CK reliance varies across conditions and models, motivating new methods to detect and evaluate lexical conflicts and guide models' lexical choices in LLM-based translation.

en