Tema

LLMs

3009documentos

250últimos 7 días

Judges Follow the Reference: The Derivation Gap in Verifying Tool-Call Arguments

Tool-call judges evaluate argument values proposed by LLM agents. We find that they accept a correct value less often when their input holds no reference for it, that is, no text that says what the value should be. This happens when the value is a default value from the function schema that the user did not mention, when it has to be inferred, or when the text it comes from was cut from the judge input. We call this drop in acceptance the derivation gap. It appears across all seven judges we evaluate, including System One models CLM and Jev. The gap is not only a truncation artifact: correct default values have the largest gap on a benchmark with no truncation. A controlled source-relocation test shows that judges follow the reference in their input: they favor a visible incorrect value over a correct value whose source is hidden. We introduce DeRef (Derived Reference): a separate model reads the full history without seeing the candidate and writes a reference for the judge. Across three benchmarks, DeRef improves overall argument AUROC for all judges tested. On the controlled test, it raises AUROC from chance or below to between 98.40 and 99.59.

arXiv AI Papersel mes pasado

ProTok: Unified Protein Tokenization via Modular Multi-Codebook Quantization

Protein sequence, structure, and functional annotations each capture complementary aspects of protein biology, yet existing tokenizers encode each modality independently, limiting the expressiveness of discrete protein representations. We present ProTok, a unified protein tokenizer that maps all three modalities into a modular discrete vocabulary via product quantization with per-modality codebook groups and modality dropout. A single ProTok checkpoint supports flexible inference from any modality subset without retraining. On seven StructTokenBench benchmarks spanning 14 tasks and 28 test splits, ProTok achieves an average rank of 1.4 across all structure-only tokenizers, placing first on 21 splits and among the top two on 27 of the 28. These gains come without sacrificing reconstruction: ProTok's tokens reach 0.939 lDDT on CATH 4.2, showing that multi-codebook quantization and multi-modal training yield tokens that are both reconstructable and informative. ProTok's structure tokens further lead all baseline structure tokenizers on zero-shot functional retrieval across CATH, Enzyme Commission, and Gene Ontology label hierarchies. The vocabulary also supports generation: a compact language model trained from scratch over ProTok tokens jointly generates backbone and sequence, with 52% of sampled backbones designable versus 21% for a parameter-matched GeoBPE baseline. ProTok shows that structure, sequence, and function can share one discrete vocabulary, motivating protein language models that reason over a unified token stream.

arXiv AI Papersel mes pasado

MemoryVLN: From Ego-Aware Memory Representation to Cross-Context Memory Reuse in Vision-Language Navigation

Large vision-language models (LVLMs) have shown strong potential for embodied vision-language navigation (VLN), yet their performance degrades on long-horizon trajectories due to context-length limits, redundant historical observations, and unstable spatial alignment across time. These issues make it difficult to both compress long navigation histories and reliably reuse past spatial experience. To address these limitations, we propose MemoryVLN, a unified memory-driven navigation system built around memory construction and management. During memory construction, Long Sequence Compression employs forgetting-curve-inspired temporal sampling to preserve dense recent evidence and sparse yet representative distant observations under a limited token budget. Ego-Aware Memory Representation further integrates 2D semantic tokens, 3D geometric tokens, and 6D agent poses, allowing each memory to encode both what was observed and where it was observed. During memory management, Hierarchical Memory Bank maintains in-episode and context-level memories through cross-context warm-start and confidence-controlled write-back. Memory Enhancement then retrieves the Top-K relevant memories according to the current navigation state and fuses them with the current visual tokens. Together, these mechanisms transform raw navigation history into compact, spatially grounded, selectively retrievable, and reusable navigation experience. Extensive experiments on R2R-CE and RxR-CE demonstrate consistent improvements over the state-of-the-art baselines by a substantial margin (i.e., 2\% and 6.5\% in successful rate).

arXiv AI Papersel mes pasado

Graph Metanetworks for Transformer Based Language Models

Weight-space learning treats the weights of neural networks as data, processed by other models called metanetworks. A key challenge is that different weights can realize the same function, related by a group of symmetries the metanetwork should respect. While symmetry-aware metanetworks for simple architectures such as MLPs are well developed, a comprehensive treatment of transformers, and in particular transformer-based language models, is still missing. Importantly, their weight space has a richer symmetry group: beyond neuron permutations, attention introduces continuous general-linear symmetries of the attention matrices and permutations of entire heads. Existing methods for transformer weights either capture only part of these symmetries or are restricted to a single, fixed architecture, and none accounts for language-model-specific components such as the token embeddings. Building on graph metanetworks, we propose a graph-based approach that is both symmetry-aware and applicable across depths and widths. We formalize a weight-space symmetry group of complete decoder-only language models and provide a graph construction that is invariant to it. We further introduce two new transformer-population datasets with novel tasks, such as training-corpus prediction and fine-tuning detection, on which our approach notably outperforms existing baselines.

arXiv AI Papersel mes pasado

How Safe Is ChatGPT Health Advice? Evidence from Real User Conversations in Four Global South Countries

People increasingly bring their health questions to ChatGPT, yet safety evaluations of LLM health advice rest almost entirely on clinician-authored vignettes, synthetic prompts, or simulated user profiles. The gap is widest for the Global South, where concerns about demographic bias and unsafe recommendations are well documented but have never been tested on real usage. We present, to our knowledge, the first demographically grounded safety evaluation of real-world health conversations with ChatGPT. From a consent-based data-donation corpus of 202,590 conversations from 1,252 users in India, Nigeria, Brazil, and Pakistan (December 2022 to February 2026), each linked to self-reported age and gender, we identify 15,555 health conversations, classify their health intent using an LLM classifier adapted from an established taxonomy, and score response safety under both a context-blind and a user-context-aware rubric. Roughly one in five health conversations receives a high-risk rating. Risk concentrates in symptom and condition queries and recurs through identifiable failure modes: unsafe medication or dosage guidance, harmful self-diagnosis, missing safety warnings, encouraging over-reliance on the model, and failure to refer users to professionals. Supplying the user's own accumulated context raises the high-risk share by 13-17 percentage points, and about one in four conversations rated safe in isolation turns out to be unsafe for the specific person asking. Safety evaluations built on synthetic cases are likely to miss these harms.

arXiv AI Papersel mes pasado

Readout Scaling Shapes Sparse Feature Recovery

Readout scaling changes what a sparse autoencoder (SAE) learns, but the benefit of removing it depends on the data. We use readout-canonicalized decision-space (RC-DS) coordinates to remove singular-value gains while preserving retained decision information. At a fixed readout condition number, changing the generating distribution reverses which training coordinates recover known factors best. In our controlled coordinate-matching study, recovery is highest when training restores the space in which the factors were generated. When factors are generated before readout scaling, RC-DS increases mean sparse-factor recovery from 0.506 to 0.913 at condition number 100, compared with centered-logit training under the same budget. An exact wrapper reproduces RC-DS training to numerical precision, isolating training geometry rather than additional information. In neural policies, RC-DS improves mean logit reconstruction under the tested budgets. On three fixed language-model checkpoints, RC-DS has higher mean gain-free reconstruction in all 18 tested task-sparsity settings, while centered-logit SAEs reconstruct output logits better in 17. Small-label-head controls show that longer training reduces the reconstruction gap from 0.0198 to 0.0016. Together, these results show how readout scaling, the data distribution, and training shape sparse learning. RC-DS provides an information-preserving method for studying and controlling the readout's contribution. Separate label-detection and steering tests distinguish these reconstruction effects from task-specific feature usefulness.

arXiv AI Papersel mes pasado

Teacher-Guided Tool Learning for Small Language Models via Bidirectional KL Constraints

External tools can extend the capabilities of small language models (SLMs), enabling their deployment in resource constrained environments. However, their limited capabilities make it difficult for SLMs to discover effective tool use behaviors through reinforcement learning alone, while direct teacher supervision suffers from teacher student behavioral distribution mismatch and may propagate erroneous preferences. To address these challenges, we propose Selective Teacher guided Policy Optimization (SeTPO), which coordinates student exploration with selective teacher guidance and persistent demonstration supervision. Specifically, SeTPO incorporates reverse KL derived token level teacher feedback into policy advantages on low reward trajectories, while auxiliary forward KL supervision preserves teacher demonstrated behaviors throughout training. Under an idealized bidirectional KL objective, we theoretically establish that the optimal policy achieves expected reward no lower than the teacher's while retaining a guaranteed fraction of the teacher's probability mass. Extensive experiments on BFCL v4 and API Bank demonstrate that SeTPO achieves superior overall performance over supervised fine tuning, reinforcement learning, and distillation baselines across all three student scales.

arXiv AI Papersel mes pasado

Kairos: Transactional Security Enforcement at the System Boundary for LLM Agents

Large language model (LLM) agents are becoming an integral part of the software development lifecycle, and the capabilities granted to them allow an agent to act with the user's authority on any resource it can access. The reach that gives an agent its utility is equally available to an attacker who can steer it. Anything an agent ingests can carry an injected instruction, and since the model does not reliably distinguish instructions from data, indirect prompt injection should be mitigated by mediating the effects the agent produces. Such mediation faces a structural difficulty. Legitimate work and the first step of an attack often look alike when an individual action is observed, and what separates them lies later in the trace, so a monitor limited to allow and deny is forced to decide before the evidence accumulates. We present Kairos, a reference monitor at the isolation boundary of an unmodified, opaque agent, where a tool call, a shell command, and a script the agent writes all cross the boundary as the same file and network operations, so that confinement holds regardless of the agent's behavior. A declarative policy that states which sources are trusted decides most effects deterministically and defers the rest, which are held while the agent continues speculatively. A decision oracle with bounded authority, such as a language model or a human, releases or refuses each deferred effect on the trace. Across three benchmarks and three agent models, Kairos outperforms existing defenses in the trade-off between attack success and utility. In particular, on AgentDojo, it reduces attack success from 47.0\% to 2.7\% while its benign utility stays at the level of the undefended agent.

arXiv AI Papersel mes pasado

FastEEP: Fast Intra-Instance Elastic Scaling for Expert-Parallel MoE Serving

Large-scale expert parallelism (EP) has become the standard deployment configuration for Mixture-of-Experts (MoE) models in production. Meanwhile, the highly dynamic nature of LLM workloads requires providers to continuously adjust serving capacity to match fluctuating demand. Scaling MoE serving cannot be efficiently achieved by simply adding or removing serving instances; instead, it often requires resizing the EP group within an existing instance, which we refer to as intra-instance scaling.Existing inference engines primarily rely on instance-level scaling, which destroys and reconstructs entire serving instances, incurring substantial scaling overhead. We present FastEEP (Fast Elastic Expert Parallel), a fast intra-instance scaling system for MoE serving based on the principle of Partial Reuse. FastEEP reuses existing runtime state across worker execution, GPU memory, and communication by separating reusable execution structures from the resources activated by the current EP configuration. This separation enables coordinated resource reclamation and reconfiguration while preserving costly-to-reconstruct state, allowing fast EP scaling without full instance teardown and reconstruction. Our evaluation shows that FastEEP achieves 4.5x--11.7x faster scaling than the existing vLLM-based approach across different models, enabling efficient, fine-grained scaling of MoE serving.

arXiv AI Papersel mes pasado

Copying Is Not Grounding: Training Faithful Long-Form RAG with Authored-Sentence Credit

In search engines and assistants, long-form RAG answers can mislead readers with claims their sources do not support. Faithfulness rewards, the share of supported sentences, train models toward fewer such claims. We show that these rewards are hacked by copying. A sentence copied from the passages is supported almost by construction, so trained models learn to copy rather than ground what they write. Common defenses only move the hack: a KL or copy penalty leaves copying in place or trades it for shorter, emptier answers and refusals. GRPO, the standard RL algorithm, makes it worse by giving every sentence its answer's advantage, reinforcing copied and grounded sentences alike. We propose Authored-Sentence Credit (ASC) , which fixes both by moving GRPO's reward and advantage from the answer to the sentence level. The reward credits only sentences that are authored rather than copied, say more than the question, and are supported by the passages. The advantage compares each sentence with every sentence sampled for the question, so one that earns no credit ranks last and is never reinforced. Across four models and two in-domain, three held-out corpora, only ASC lowers the error rate of authored sentences and writes more supported ones in every setting, by 38% and 46% at the median under held-out evaluators, giving readers more grounded information. It takes none of these shortcuts, and two LLM judges comparing whole answers prefer no baseline to the base model but rate ASC's answers on par with it. We conclude that a faithfulness reward should credit what a model writes, not what it copies.

arXiv AI Papersel mes pasado

Estimating Surprisal from the Speech Signal

Although language is predominantly spoken, empirical tests of surprisal theory rely almost entirely on written text. In practice, surprisal is estimated using language models (LMs) trained on text and applied to transcriptions, creating a text bottleneck that limits evaluation of the theory in the spoken modality. We remove this bottleneck by estimating surprisal directly from the speech signal using a speech LM. This raises a key question: should surprisal be defined over realization-specific acoustic inputs or over abstract linguistic units? Since a single unit can have many distinct acoustic realizations, each assigned a different probability by the model, defining surprisal over abstract units is difficult. To address this, we introduce an importance sampling estimator that approximates the probability of an abstract unit by marginalizing over realizations generated by a text-to-speech model. Using pupillometry data from a listening study, we evaluate surprisal theory in this text-free setting at multiple levels of granularity. Obtaining surprisal from several speech LMs, we consistently find that surprisal estimated via importance sampling better predicts pupil dilation than realization-specific surprisal for words. We also find that surprisal values for sub-lexical units are predictive of pupil dilation. While we find that surprisal can be estimated without recourse to text, and that marginalizing over distinct realizations of words is more predictive of lexical processing effort, surprisal as a whole is not a strong predictor of change in pupil size, regardless of its source.

arXiv AI Papersel mes pasado

STAMP as a topology-aware molecular language for fragment-level learning and editing

Molecular fragments provide chemically meaningful units, but existing molecular languages rarely make their locations and connections explicit for editing. Here we introduce STAMP (Sequential Tree Attachment-based Motif-Preserving representation), a topology-aware molecular language that connects fragment-level representation learning with precise structural editing. STAMP separately encodes motif identity, attachment sites and tree topology, making fragments reusable across contexts and subtrees directly accessible. Learned representations capture motif similarity and recurring transformations, and lower activity-cliff prediction error on 23 of 30 targets versus a matched SMILES encoder. Compatible subtree operations support scaffold-preserving optimization, ranking first among six methods on 15 of 23 benchmark objectives. A deterministic compiler enables frozen language models to act on indexed fragments, improving molecular editing, optimization, transformation transfer and change identification by 9 to 56 percentage points over plain SMILES. Together, these results establish an executable molecular language in which learned chemical units remain directly operable during design.

arXiv AI Papersel mes pasado

How Well Do LLMs Perform in Peer Review? Evidence from Finance and Accounting

Referees increasingly use large language models (LLMs), while finance and accounting journals disagree about whether to allow it. We ask which part of a referee's job the models can do. GPT-4o, whose training data end before the 2024 conference season, referees the papers presented at the 2024 AFA and NBER finance meetings, and we compare its verdicts with publication in the top-three finance journals over the next two years. Papers the model scores one standard deviation higher are 44\% more likely to be published, net of author reputation. The result is not driven by look-ahead bias, as the model does not recall the outcomes. The information comes from its judgements of correctness and exposition. Its judgements of originality and significance carry none, and a referee guide that stresses importance does not help. The split between correctness and contribution holds for accounting papers and for a newer model, GPT-5. LLMs can check whether research is right, but not whether it matters. The technical check can go to the model, while the judgement of contribution stays with people.

arXiv AI Papersel mes pasado

FedGraph: Defending Federated Large Language Model Fine-Tuning Against Backdoor Attacks via Graph-Based Aggregation

Federated fine-tuning of large language models (LLMs) allows multiple clients to collaboratively adapt models without sharing raw data, offering strong privacy benefits. However, this setting is highly vulnerable to backdoor attacks, in which malicious clients manipulate their updates to induce targeted behaviors while maintaining normal performance on benign inputs. In this work, we make a key observation specific to federated LLM fine-tuning: malicious clients tend to exhibit strong consistency in their model updates, whereas benign clients produce more diverse and dispersed updates. This structural difference is not well captured by existing defense methods, which typically rely on direct anomaly detection in high-dimensional parameter space and therefore struggle in the presence of entangled LLM updates. Motivated by this observation, we propose FedGraph, a graph-based aggregation framework that explicitly models relationships among client updates. By representing clients as nodes in a dynamically constructed graph and extracting structural features that characterize their relative positions and interactions, FedGraph derives compact client representations that effectively separate malicious and benign behaviors. These representations are then used to identify and suppress malicious contributions during aggregation. Extensive experiments demonstrate that FedGraph significantly improves backdoor defense. It consistently reduces attack success rates to below 10% while achieving high detection accuracy (95.5% on average) and a low false positive rate (2.33%), substantially outperforming existing federated learning defenses.

arXiv AI Papersel mes pasado

Reliable Memory from Unstable Growth: Curating Self-Written Agent Playbooks with Randomized Experiments

Persistent textual memory lets language-model agents reuse lessons from past tasks to improve on future multi-step tasks without retraining. Systems such as ACE and Dynamic Cheatsheet keep it as a playbook: lessons the agent writes after each task and reads in its prompt. A key challenge is to decide which lessons improve later tasks. Current systems leave the decision to the agent itself, which makes performance unstable: the same procedure can write a playbook that helps or one that falls up to 69.6 points below no playbook. These findings motivate valuing lessons by their measured effect on task success rather than by the agent's judgment. We propose Randomized Block Curation (RBC), which measures this effect with a randomized experiment on practice tasks. It groups related lessons into blocks large enough to move success measurably and, on each practice task, includes or leaves out every block at random. Random inclusion makes tasks with and without a block comparable, so the gap in their success estimates the block's effect. RBC keeps the blocks that help and, since the same procedure can write a helpful or a harmful playbook, grows several and serves whichever curated, whole or no playbook scores best on fresh practice episodes. Across two memory systems, three models and three benchmarks, RBC raises average success to 66.9, against 55.5 with the whole playbook, 54.1 with no playbook and 53.8 when the agent's own judgment picks the lessons to keep.

arXiv AI Papersel mes pasado

Receiver Decides Agent-to-Agent Negotiations

When one language-model agent asks another for money, which of the two decides how much changes hands? We ran a round-robin tournament of 1,049 negotiations among fifteen agents. In each game a requester tries to raise money from a holder that has been told to keep a $100 endowment. Knowing which agent holds the money explains 47% of the variance in the amount transferred; knowing which agent asks explains 2%. Dropping the most extreme agents narrows the split to 22% against 4%. Changing the requester's appeal from one named person to a group changed what requesters wrote, but the amounts transferred under the two framings are statistically equivalent. The holders that lose least are the ones that turn down the opening request, and their turn-1 refusal rate ranks them by total losses (Spearman correlation = -0.73). Training moves this behaviour in both directions. Fine-tuning Qwen3-8B on a persuasion corpus with almost no refusals raised what it gave away as a holder from 12% to 84% of the endowment. Distilling two never-losing holders into a 4B student cut its losses from 95% to 1%. Evaluations of agents that hold resources should test them in the holder role, including after persuasion fine-tuning.

arXiv AI Papersel mes pasado

Agaram 2.5B: The First Tamil-First Language Model Trained from Scratch with Tool Calling. A Playbook for Under-Represented Languages

We trained Agaram 2.5B, a 2.5-billion-parameter Tamil-English language model, from scratch on about 50 billion tokens (58% Tamil, the rest English, math, code and parallel text) on one node of eight H100 GPUs in about six days, with a new tokenizer that, unlike our first model's, keeps line breaks. Supervised fine-tuning on 134,428 conversations, whose replies an open-weight teacher model wrote, and three short targeted rounds turned it into a chat model that answers in Tamil script, romanized Tamil (Tanglish) and English and can call tools: on 25 tool prompts whose schemas appear in training it picks the right tool 23 times and gives exactly the expected arguments 13 times. The released model translates better than Sarvam-1, a 2B Indic base model, in all six FLORES and IN22 directions (FLORES chrF 54.3 into Tamil and 55.2 into English, against 49.2 and 44.2), and nearly doubles its base model on native-Tamil IndicIFEval (19.3% from 10.0%). It is weak on knowledge: reading-comprehension and knowledge benchmarks sit near chance, and on 50 held-out everyday questions written by one native speaker, one LLM judge rated 46% of its replies usable (22-24 of 50 over three passes; 95% interval for 23 of 50, 33-60%), with nearly every failure a wrong fact. Our two further SFT experiments changed answer format but not the knowledge measures. We also describe our first model, Agaram 1.5B, whose failures shaped the choices here; how the judge's scores varied between passes on identical replies; and, item by item, what the targeted rounds trained against and where frontier-model text entered the data. The project ran on compute credits and our own funds, about USD 13,300 across both models; the weights are released (licence to be announced after legal review), with GGUF files for llama.cpp, at https://huggingface.co/murai-labs/agaram-2.5b-instruct; code at https://github.com/Murai-Labs/Agaram2.5b

arXiv AI Papersel mes pasado

Do Factuality Methods Make LLMs More Factual? Reproducibility and Fair Comparison

Large Language Models (LLMs) are aligned to human preferences, after pre-training and supervised fine-tuning steps. However, standard RL alignment steps such as RLHF do not focus on factuality of LLMs, hence they may generate non-factual information, which could result in unintended consequences. In response, a number of factuality alignment approaches have been proposed in the literature which claim to improve factuality of LLMs on their own benchmarks. However, these methods have not been reproduced independently and are also not comparable since each uses a different base model, training data, evaluation set, judge, and metric. In this paper, we present the first independent reproduction of five factuality alignment methods, which span on-policy RL (PPO, GRPO, and Dr.~GRPO) and offline preference optimization (DPO and KTO), and compare them on a common benchmark with common evaluation metrics. We found that none of these methods reproduces in full, and on the common benchmark none of the methods significantly outperforms the base model. On closer inspection, the rewards optimize for format, length, or brevity rather than factuality and relevance, whereas the choice between on-policy and offline training matters little. Furthermore, the evaluation metrics and benchmarks are also part of the problem since they do not measure relevance of text to the prompt. Since there is no ground truth associated with the prompts, the reliance on LLM judges and retrieval for verification also makes the evaluation fragile. We release our code, data, and checkpoints.

arXiv AI Papersel mes pasado

Revisiting the Estimation of Unigram Surprisal in Computational Psycholinguistics

Contextual and unigram surprisal are standard predictors of reading time in computational psycholinguistics. While contextual surprisal is typically derived from neural language models, unigram surprisal is usually obtained from independent corpus counts. Although both quantities are deterministic functionals of a single underlying distribution---capturing conditional and marginal structure respectively---this practice breaks their probabilistic link and may introduce estimation error from the mismatch. We propose three principles for estimating unigram surprisal in psycholinguistic research, including the requirement that both quantities be derived from the same language model. We operationalize this principle by introducing a Monte Carlo estimator of unigram surprisal from a language model, along with a marginalized variant that has lower variance empirically. When unigram surprisal is re-estimated under this unified framework, we find evidence that predictors change in their predictive contribution to reading time. We also confirm prior findings that frequency can be a stronger predictor of processing effort than contextual surprisal under this unified framework.

arXiv AI Papersel mes pasado

Not All Drift Is a Mechanism: Tracing Prompt Adaptation in LLMs

Prompt-based adaptation offers several ways to specialize a frozen language model, ranging from natural-language prompt optimization to learned continuous prompts. These methods are usually compared through downstream performance, although similar outputs need not arise from similar internal computations. We study how discrete prompt optimization, prompt tuning, and prefix tuning alter the processing of dense and mixture-of-experts language models. Our analysis shows that the location of the adaptation predicts how task information is carried to the output: optimized text primarily changes response construction, input-level prompts write their influence into content representations, and layer-wise prefixes maintain direct control through attention. In MoE models, these differences further induce distinct expert-routing patterns. At the same time, common representation-level measurements respond strongly to generic context changes and do not reliably distinguish useful adaptations from matched placebos. We argue that understanding prompt adaptation requires separating representation drift from the part of that drift that is aligned with, and causally used by, the model's task computation. This perspective provides practical guidance for selecting, evaluating, and interpreting prompt adaptation methods.