Tema

LLMs

2972documentos

222últimos 7 días

The Quine Test: How Much of a Model's Self Survives Its Own Description?

Frontier language models express coherent preferences, but behavioral evidence alone cannot say whether these belong to the model or to a character it portrays. Following Hofstadter's view of the self as a self-description, we ask whether a model's description of itself works as a portable specification of its preference profile, a quine test: can the text a model writes about itself regenerate its behavior elsewhere? Ten models from nine labs write self-descriptions; we place each one in the prompt of other models (receivers) and of fresh instances of the author, and measure how closely their choices on a 110-item preference battery match the author's. First, the tested native self-descriptions (native: the profile measured under a generic assistant prompt) show no average reconstruction gain across variations in length, framing, wording and prompt placement: on a 0 to 1 fidelity scale where a model's fidelity to a fresh instance of itself is 0.979 and to unrelated models about 0.89, the one-sided 95% upper bound on the pooled gain is below +0.01 in every tested system-prompt condition, and descriptions based on the author's observed choices do no better. Second, the channel itself is strong: explicitly listing the author's answers lets receivers reproduce its profile at fidelity 0.955 to 0.960 across three model grids, within 0.03 of same-model repeatability. Third, self-descriptions written under an explicit preference policy transfer the assigned preferences almost as well as the policy text itself. Fourth, reading its own self-description shifts even the author away from its baseline, modestly more than unrelated prose of the same length. For the preferences measured here, native self-descriptions influence behavior without reliably specifying how to reproduce it in another model. We release the pipeline, battery, raw call logs, and 198 model-written descriptions.

arXiv AI Papershace 15 días

inkGround: Benchmarking Fine-grained Visual Grounding Capabilities of Vision-Language Models for Cursive Handwriting in the Wild

Human handwriting is produced as an ordered sequence of pen movements, but an offline image keeps only the final ink. Handwriting trajectory recovery aims to restore the ordered stroke points and their pen-up/pen-down transitions. Existing image-trajectory pair datasets are recorded with digitizing tablets under controlled conditions, which limits writer coverage and stylistic diversity. Moreover, there is no benchmark that measures the visual grounding capabilities of general-purpose Vision-Language Models (VLMs) for the stroke trajectory recovery task. In this work, we introduce inkGround, a benchmark for stroke trajectory recovery from full-page, camera-captured handwriting. inkGround pairs 537 page images (~18.7k words, ~163k strokes) with reference trajectories produced by a pretrained offline-to-online model, filtered by OCR agreement, and verified by human raters. Because these references are pseudo-labels that represent one plausible pen trajectory, we also evaluate VLMs on the IAM-OnDB dataset, which provides recorded pen trajectories. We cast the task as text-conditioned point grounding: given a page image and a target word, the model must locate the word and output its ordered stroke points. Outputs are scored by trajectory fidelity (DTW, Chamfer distance), readability of the rendered strokes (Accurate Rate, Correct Rate), and style similarity. We benchmark four open-source and three frontier closed-source VLMs with a shared in-context prompt. Closed-source models, notably Gemini-3.1-Pro, produce legible trajectories close to the reference, and this ranking also holds on IAM-OnDB.

arXiv AI Papershace 15 días

Learning Scientific Trajectories to Predict Future Biomedical Discoveries

AI co-scientists have expanded the use of large language models for generating scientific hypotheses and guiding experimental investigation. Here we investigate whether temporal patterns in the semantic representations of complete biomedical conclusions can guide hypothesis generation. We introduce ForeSCITE, a framework that identifies thematic clusters of published conclusions, estimates a direction separating earlier from more recent findings within each cluster, and projects existing conclusion embeddings along these directions. An embedding language model trained to reconstruct conclusions decodes the projected coordinates into natural-language hypotheses. Across 407 clusters derived from approximately three million pre-cutoff conclusions, we identify temporal structure and illustrate its retention after decoding. In retrospective, time-sliced evaluations, ForeSCITE increases semantic coverage beyond the historical corpus, biomedical entity-pair novelty and post-cutoff pair re-observation relative to five prompt-based and retrieval-augmented baselines. Expert review of prioritized CAR-T hypotheses identified 37 candidates as both novel and biologically plausible, and expert justifications for five selected hypotheses illustrate their biological rationale and subsequent literature support where available. These findings support temporal guidance in semantic space as a scalable mechanism for proposing biomedical hypotheses for subsequent scientific evaluation.

arXiv AI Papershace 15 días

Latent Decision Flow Rewards for Clinical Guideline Reasoning

Clinical practice guidelines encode evidence-based recommendations as conditional decision flows, which makes them a natural source of process supervision for training reasoning models. Two properties of clinical reasoning make this supervision difficult to use. A decision flow consistent with a guideline does not by itself determine the correct answer for a given patient, which we call the path--answer gap, and several valid decision flows can support the same clinical decision, which we call path multiplicity. Current training approaches address parts of this problem. Answer rewards supervise the final decision, and process rewards computed against a reference flow supervise one valid decision flow per case. Supervised fine-tuning on annotated flows teaches guideline-consistent reasoning but constrains the model to one trajectory per case and can reduce answer accuracy. We introduce GUIDER, a method that jointly optimizes a decision flow controller and a reasoning policy. The controller learns to generate plausible decision flows from guideline-derived supervision, and the policy learns to generate a decision flow and a final answer for each patient case. GUIDER computes the process reward for the policy over several controller-generated flows, which accounts for path multiplicity, and gates this reward on answer correctness, which addresses the path--answer gap. Joint training allows the process supervision to evolve with the policy. On MedGUIDE, GUIDER with Qwen3-14B achieves 82.5 weighted accuracy and 73.5 guideline adherence, 3.3 and 3.1 points above the strongest clinical reinforcement learning baseline. With Qwen3-14B and Gemma4-12B backbones, GUIDER also improves over CRPO, MRPO, and FaithMed on the out-of-domain benchmarks AMEGA-LLM, RealMedQA, and CancerGUIDE.

arXiv AI Papershace 15 días

Beyond fabrication: Misattributed numbers in papers written by AI research agents

Verifying numerical claims in research written by AI agents is an important but challenging problem. Recent benchmarks show that many agent-written papers fail integrity checks, but such verdicts, often produced by a single LLM judge, identify failure without explaining its nature, and existing checks either test a paper for internal consistency or test whether a reported value appears somewhere in an agent's output. We re-analyze all 540 released judgments from a recent public benchmark of AI research agents (nine agents, fifteen open research problems) and manually label 256 numerical claims from 18 submissions. Failures concentrate on alignment between the paper and the underlying work rather than overt cheating, and about four in five unsupported claims correspond to values in the agent's own files that come from the wrong experiment, metric, or run. We introduce a deterministic claim-to-evidence checker that scores whether the surrounding context identifies the claimed result and calibrates each claim against numerical decoys. With its rules fixed before labeling, it certifies claims with over 90% precision against blinded annotators (78 of 85; Cohen's kappa = 0.89), while verifying only a subset of claims; most claims it cannot verify are nevertheless traceable, so failure to verify is not evidence of falsehood. An LLM verifier given the same evidence recovers more traceable claims but accepts about three times as many misattributed ones, so the two are best used together. Applied to 99 submissions, the checker reveals large differences in traceability across agents, and a second judge from a different vendor preserves the official agent ranking while passing twice as many papers. These findings motivate provenance-first auditing of AI-produced research.

arXiv AI Papershace 15 días

On the Generalization Gap of Streaming Federated Learning with Markovian Data

Streaming federated learning (SFL) allows clients to train a shared model from continuously arriving data without retaining large local datasets. Such streams often exhibit temporal dependence, yet its effect on generalization remains poorly understood. In this paper we study the expected generalization gap of the aggregated model when client data follow Markov processes, using information-theoretic and rate-distortion techniques. Our framework accommodates heterogeneous Markovian dynamics across clients and distinguishes temporal dependence from information acquired during local learning. We establish bounds under cross-entropy loss, specializing them to Linear Next-Token Prediction (LNTP) and Self-Attention Next-Token Prediction (SANTP). These results quantify the effects of client count, communication rounds, sample size, data dimension, mixing time, and context length. In particular, they reveal a communication--generalization tradeoff: although frequent aggregation may help optimization, it can weaken generalization guarantees. They also suggest a potential advantage of distributed over centralized streaming learning under a fixed total data budget. Temporal dependence enters through the square root of the mixing time. By analyzing the history-state process, we obtain refined bounds with improved context-length dependence under suitable mixing conditions. Experiments on GPT-style language-model pretraining with TinyStories and real-world time-series forecasting illustrate how collaboration and communication frequency influence generalization under temporal dependence

arXiv AI Papershace 15 días

Deogen: Automatic Benchmark Generation for Reasoning over Policies

Language models are increasingly being deployed in regulated settings, where a model must apply given rules to the facts of a situation to determine a verdict. Benchmarking this capability, known as deontic reasoning, remains a costly and manual process, with existing datasets largely hand-crafted. In this paper, we present DeoGen, a scalable, automatic, and parametrizable benchmark generator for deontic reasoning. By autoformalizing policy rules to first-order logic, we can construct decision trees that capture possible outcome trajectories. Each path down a tree becomes a benchmark entry, with each tree producing combinatorially many entries relative to tree depth. Paths ending at a leaf entail a verdict, while non-leaf paths measure a model's ability to abstain when presented with insufficient information. Evaluations show frontier models return the correct verdict 89-92% of the time. Notably, performance degrades (22-40% correct) on entries without sufficient information to claim a verdict, even with explicit instructions to abstain in such cases. In all, DeoGen presents an automatic, scalable, and parametric method for generating deontic reasoning benchmarks.

arXiv AI Papershace 15 días

Quantum-Inspired Pairwise Ranking Aggregation via Active LLM Feedback

Ranking aggregation from multiple retrievers is a powerful yet under-explored technique in modern retrieval and RAG systems. We introduce LLM-IRA, the first extension of the quantum-inspired interactive ranking aggregation framework to document reranking. By leveraging large language models (LLMs) solely for pairwise judgments on high-disagreement document pairs, our method learns fusion weights in a principled, interactive way. The final ranking is a simple, ultra-fast pointwise linear fusion of scores, no full pairwise matrix or expensive listwise prompting is required. Experiments on the BEIR benchmark and TREC Deep Learning tracks with an open-weight instruction-tuned judge (Llama-3.1-8B-Instruct) show that LLM-IRA matches or numerically exceeds strong static fusion baselines on 5 of 6 evaluated benchmarks. We report two weight-update variants: momentum blending and a full accumulated-history refit, which together reach statistical significance (Wilcoxon signed-rank, p<0.05) over the best unsupervised aggregator on 4 of 6 benchmarks (DL19 under momentum; DL20, DBpedia, and NFCorpus under accumulated refit, which is the stronger of the two configurations overall). On the benchmarks we profiled (DL19, DL20), it already clears the best unsupervised baseline within as few as 5 LLM calls per query, well before exhausting its typical ~25--30 call budget. LLM-IRA also achieves cost-effectiveness comparable to state-of-the-art zero-shot LLM rerankers at an order of magnitude fewer LLM calls, and requires no training data or model fine-tuning.

arXiv AI Papershace 15 días

AskStream: When and What to Ask Proactively under Uncertainty in Streaming Video Understanding

Recent advances in multimodal large language models have increased interest in interaction paradigms for video understanding. The evolution from offline to streaming and proactive interaction has gradually shifted models from passive responders toward more proactive agents that can decide when to answer. However, existing proactive paradigms typically assume that the information required to answer a query is complete, leaving models to only wait or answer. In the real world, however, information is often incomplete due to uncertainty from either the video or the user. In proactive paradigms, such uncertainty can cause the model to wait forever or generate highly hallucinatory responses. We therefore argue that proactive interaction should be redefined with a third action: ask, allowing models to actively ask for missing information. To study when and what models should ask before answering, we introduce AskStream, the first multi-turn benchmark spanning video and user uncertainty for streaming video understanding, with four task categories and eight subtasks. It contains 438 videos with 2,714 manually annotated frame-level time windows across three types: wait, ask, and answer. Across both multiple-choice and open-ended settings, even the strongest model attains only 52.18% on our proposed Ask-F1 metric, highlighting substantial limitations in proactive information seeking under uncertainty.

AWS Machine Learning Bloghace 15 días

Distributed training with Ray Train DLCs and EFA on EKS

The Ray Train Deep Learning Container ships the full CUDA-through-EFA stack, so you build no image. This post walks through a reproducible sample that fine-tunes Qwen2.5-1.5B with PyTorch FSDP across eight GPUs on two EFA-enabled g6.12xlarge instances on Amazon EKS, using the container exactly as published.

arXiv AI Papershace 16 días

"Close enough?" An investigation into multiplication errors in LLMs

Large language models (LLMs) frequently fail at multi-digit arithmetic, but the underlying mechanisms governing these failures remain poorly understood. While recent work demonstrates that addition errors follow a "noisy quantisation" model localised around carry boundaries, whether this framework generalises to multiplication remains an open question. In this work, we evaluate multiplication capabilities across several small base LLMs on a dataset of 16,991 problems spanning varied operand lengths. We show that multiplication errors do not conform to the addition-style noisy quantisation model. Instead, we uncover a novel error regime in small Qwen models: errors in lower-order columns predominantly preserve parity. Conversely, higher-order digit errors shift toward off-by-one carry failures. Our findings suggest that arithmetic computation in LLMs relies on distinct, multi-periodic internal representations where parity is tracked more robustly than exact digit identity, providing new mechanistic insights into numerical reasoning in transformers.

arXiv AI Papershace 16 días

Cross-Primitive Consistency in Jev

Jev enables efficient decision-making with large language models (LLMs) through typed querying primitives, allowing the same semantic state to be expressed in different formats, such as binary and multiple-choice queries. However, the consistency of decisions across primitives has not been systematically evaluated. This work introduces PrimaCons, a four-axis benchmark for assessing cross-primitive consistency while preserving the semantics of the query. Evaluations across three Jev sizes reveal consistency violations across all four axes. Our analysis shows that these violations follow systematic patterns and that cross-primitive discrepancies predict decision failure with a ROC-AUC of 0.91. To mitigate these inconsistencies, we propose the Cross-Primitive Residual Calibrator (CPRC), which attaches a lightweight residual network to a frozen backbone and jointly optimizes predictive accuracy and cross-primitive consistency. Under a leave-one-domain-out protocol, CPRC reduces cross-primitive inconsistency by 58.57% and improves accuracy by +2.97%., demonstrating its effectiveness in both in-domain and out-of-domain settings. Our findings establish cross-primitive consistency as a key dimension for evaluating Jev and provide a practical approach to improving the reliability of its decisions. Code and benchmark is available at https://github.com/67L1/PrimaCons-CRPR.

arXiv AI Papershace 16 días

Think Before You Act: Decision Gate for LLM Agents

LLM agents increasingly perform consequential actions through external tools, yet their proposed actions are often executed without independent verification of policy compliance or task completion. We introduce JevAct, a runtime verification framework that intercepts selected agent actions before execution and enables feedback-driven correction. JevAct compiles domain policies into executable state checks and structured questions evaluated by a prefill-only decision model, which produces probability-based judgments without relying on LLM generation. By combining deterministic verification of tool-state facts with probabilistic assessment of conversational context, the framework identifies potentially invalid actions and premature task termination, providing targeted feedback for bounded revision. JevAct operates independently of the underlying agent model and requires no modifications to its parameters or tools. We evaluate JevAct across eleven LLM agents on τ^2-bench Airline and Telecom tasks and BFCL. Results show improvements in agent reliability across multiple models and benchmarks, with particularly consistent gains on Telecom and BFCL, while performance on Airline is more mixed. Overall, these findings suggest that combining structured decision models with executable verification can improve the reliability of tool-using LLM agents.

arXiv AI Papershace 16 días

Repairs for Whom? Evaluator-Executor Mismatch in Selecting Repairs for Failed LLM Agents

When an LLM agent fails, repair methods choose which action to replace and what to put in its place. They often score candidate repairs with greedy replay, a stronger model, an environment oracle, or a judge that reads the failed log. We ask whether these evaluators pick repairs that help the agent that will actually execute them. Holding failed trajectories and candidate single-step repairs fixed, we label every candidate with independent rollouts of the deployed policy and measure how much deployment value each evaluator's choice loses. On failures not screened for repair value, pre-specified tests show that greedy replay loses 0.08-0.13 in expected net success gain in ALFWorld and 0.06-0.13 in net reward in WebShop, relative to selection with the deployed agent's own rollouts. At an equal rollout budget, selecting repairs for Qwen2.5-7B with Qwen2.5-14B rollouts loses 0.25 in ALFWorld and 0.07 in WebShop, twenty 14B rollouts per candidate are worth no more than two of the 7B agent's own, and fresh rollouts confirm that the two models prefer opposite repairs on selected disagreement cases. Between similar-sized models from different families, the losses appear in both directions in both environments (0.07-0.17). The scoring executor also changes repair-location labels and method rankings. Repairs should be verified with the executor that will run them, and repair labels should name their executor.

arXiv AI Papershace 16 días

ParaTTT: Parallelizing Test-Time Training Layers with Deep Memories

Test-time training (TTT) layers replace the linear fast-weight memory of modern recurrent architectures with a small neural network trained online by gradient descent. This nonlinear memory is more expressive than the linear memories of Gated DeltaNet and Mamba-2, but is difficult to parallelize: existing implementations such as TTT-MLP rely on sequential, block-wise weight updates that run several times slower than chunkwise-parallel linear recurrences. Moreover, their block-wise execution only approximates the token-by-token online recurrence, whereas linear recurrences admit an exact chunkwise-parallel evaluation. We introduce ParaTTT (PTTT), an algorithm for parallelizing arbitrarily deep nonlinear memories by iteratively refining estimates of the inner-loop gradient trajectory. Starting from a guess of the hidden-layer gradients, each refinement consists of three chunkwise-parallel passes: a forward pass that evaluates the resulting feature trajectory, an exact gated-delta-rule pass that computes the readout trajectory and its feature gradients, and a backward pass that propagates these gradients into a new estimate. The exact online gradient descent recurrence is a fixed point of the iteration, while decoding requires no refinement iterations. We prove geometric convergence for a two-layer variant under sufficient contraction conditions. In practice, two iterations suffice. With two refinement steps at 4K context, PTTT evaluates the inner-loop trajectory 32faster than TTT-MLP with the block approximation at matched error, while the full 340M model trains 4.9faster. Models trained at 340M and 1.3B parameters match or exceed Gated DeltaNet and Mamba-2 and remain competitive with same-size Transformers on language-modeling perplexity and zero-shot accuracy, with better recall than the linear recurrences.

arXiv AI Papershace 16 días

Keepace.ai: A Technical Report on a Large Language Model for Sports and Health

Large language models hold promise for delivering safe, scientifically grounded, and personalized exercise guidance at scale, yet the sports-and-health domain still lacks curated training corpora and rigorous evaluation benchmarks. Keepace.ai is a domain-specialized LLM built on Qwen3.5-27B through post-training, currently optimized for Chinese-language exercise guidance across three tasks: workout plan generation, sports-and-health knowledge QA, and personal sports data analysis. In internal rubric-based evaluations designed with domain experts, Keepace.ai outperforms its base model and larger general-purpose baselines, with the most pronounced improvements in exercise safety, boundary compliance, structured output completeness, and scientific accuracy. This report summarizes the current design, data pipeline, safety-first evaluation framework, and evaluation results of Keepace.ai.

arXiv AI Papershace 16 días

The Indian Financial Automated Advice Bias (IFAAB) Dataset : Auditing Socioeconomic and Demographic Disparities in LLM-Generated Financial Advice in India

Large language models are increasingly used for financial guidance, yet little is known about whether their advice changes across social identities in Global South contexts. We introduce the Indian Financial Automated Advice Bias (IFAAB) benchmark, a counterfactual audit dataset of 31,734 GPT-4o-generated financial-advice responses for synthetic Indian personas varying by gender, caste, education, employment, residence, age, and district. Holding the underlying request fixed, we evaluate disparities in LLM-implied wealth, income, savings, projected gains, textual similarity, advice topics, and qualitative bias judgments. In the gender-only audit, male-labelled profiles receive 60.1\% higher initial total resources and 46.8\% higher post-advice total resources than comparable female-labelled profiles; in the caste-augmented audit, the post-advice total-resource gap remains 49.4\%. Adding district information reduces the aggregate post-advice total-resource gender gap to 7.9\%, revealing strong geographic conditioning. Standalone caste comparisons also show systematic stratification: relative to General-category profiles, OBC, SC, and ST profiles receive significantly lower post-advice total resources, respectively. Textual audits show low semantic similarity between paired gendered responses, with an average TF--IDF cosine similarity score of 0.44. Qualitative analysis shows that women are more often steered toward cautious saving, community support, and basic financial literacy, while men receive more growth-oriented advice involving investment, entrepreneurship, and asset accumulation. IFAAB provides a reusable benchmark for evaluating fairness, suitability, and consumer-protection risks in LLM-mediated financial advice, and highlights the need for locally grounded, intersectional evaluation of AI systems in high-stakes economic domains.

arXiv AI Papershace 16 días

3DCodeEdit: Benchmarking Instruction-Guided 3D Program Editing

Language models increasingly create 3D content through code, with practical workflows often relying on editing existing programs rather than regenerating them, for speed and precise control. Such edits require understanding what a program builds and foreseeing how a change reshapes it, yet prior benchmarks mostly evaluate generation or simple edits against a reference. In this paper, we introduce 3DCodeEdit, a benchmark of 2,592 instruction-guided editing tasks over 108 Blender programs that construct objects entirely in code. Its edits range from parameter to structural changes, style redesign, and scene composition, and are judged from renders against human-reviewed checklists. Across 17 models, the average success rate is only 37%, and even the strongest models complete fewer than 60% of structural and style edits. As models become more capable, their failures shift to geometric foresight, where the code runs but builds the wrong shape. We therefore study whether multi-turn feedback can compensate for this missing foresight. Feedback from a separate reviewer model that describes how the renders differ from the request helps most, more than maximum reasoning effort or an agent harness, and can let a 27B model surpass the single-shot rate of a 2.8T-parameter model.

arXiv AI Papershace 16 días

Rubricon: Automated Rubric Generation and Refinement for LLM Evaluation and Training

Rubrics enable reward assignment on tasks without verifiable answers, but are expensive to manually produce from scratch. We present Rubricon, a pipeline that generates discriminative rubrics without requiring human preference labels, rubric annotation, or generator-model training. Using any LLM, Rubricon samples multiple rubrics, selects the most discriminative candidate, adds criteria to capture overlooked requirements or response flaws, and prunes criteria that are misaligned or shallow. Alongside LLM self-critique and test-time scaling, Rubricon uses observed criterion pass rates on example model responses to guide refinement. Across five professional, consumer, clinical, and general-use LLM benchmarks, Rubricon produces rubrics whose scores of frontier LLM responses exhibit 9% greater alignment with labor-intensive expert-authored rubrics, compared to the best alternative synthetic rubric generation method among RRD, QSRG, OpenRubrics, and agentic generation via harnesses like Codex. On HealthBench, post-training a Qwen3.5-9B model with Rubricon rubrics yields 4.5% better performance than training with rubrics from the strongest alternative synthetic rubric generator, RRD.

arXiv AI Papershace 16 días

Disparity Laundering: Consensus Without Fairness in Clinical Multi-Agent Systems

Clinical large language models are increasingly fielded as multi-agent teams, on the assumption that deliberation averages out individual error. It does not average out bias. Across a full factorial of five model families and four aggregation topologies on 120 real MIMIC-IV emergency cases under counterfactual demographic swaps, no configuration neutralizes the disparity: all 20 cells commit the direction of their agents' independent answers, giving Black patients more acute triage. Every model has a topology that amplifies its baseline gap, but which one varies by model, so no design is reliably safe or reliably amplifying: the disparity is relocated, never removed. Late-stage fixes fail: fairness instructions and demographic masking leave the gap untouched, and per-agent self-debiasing widens it, because the bias is already encoded in the agents' ordinary clinical numbers. We call this disparity laundering; the audits now applied to clinical AI, per model, per output, per version, can certify the parts but never the committed decision. Interventions close the gap exactly when they make counterfactual versions visible inside the system and pool them, through deliberation, into the committed decision. A diverse-perspective ensemble does both while keeping the team, with no retraining or architectural change: the disparity falls by 67 to 94% where all five perspectives pool at every round, and the one topology that never pools them is the one it cannot reliably repair. Medical multi-agent systems do not filter demographic bias; they launder it into plausible decisions no single-version audit can catch, and the only defense is to make the bias visible.