Sección

Investigación

Papers y preprints de los feeds de investigación que seguimos.

6415documentos

arXiv AI Papershace 14 días

Same Question, Different Answers: Auditing Persona-Dependent Variation in Six LLMs on Smoking-Cessation Advice

Hundreds of millions of people now ask large language model (LLM) chatbots for health advice, which raises a question that is simple to pose and hard to answer: does everyone get the same advice? We audit six widely deployed LLMs from four providers on smoking cessation, a domain with codified clini- cal guidance and substantial public-health stakes, by putting the same four cessation questions to twelve user personas that differ in life stage, self-described expertise, and smoking history, for a total of 2,880 responses. What the models say depends on who appears to be asking. Every model named FDA-approved prescription cessation medications less often to teen personas than to adult personas, by 32 to 67 percentage points, and in five of six models a persona presenting health- care expertise received fewer facts than a persona presenting none. We then evaluate one deployable intervention: a short system prompt specifying a factual floor, the set of core med- ical facts every user should receive when they are relevant, while leaving tone, vocabulary, and framing free to adapt. The prompt narrows persona differences in every model, but its effect on overall fact coverage varies in both magnitude and direction across models, so a prompt validated on one system cannot be assumed to transfer to another. Two cautions frame these results. Equal mention rates are not always the correct clinical target, because prescription cessation pharmacother- apy is not recommended for minors, so a defensible factual floor must be age-aware rather than uniform. And because our personas vary on several attributes at once, we report scenario- level differences rather than isolated effects of age or expertise. We release all 2,880 responses and the analysis code.

en

arXiv AI Papershace 14 días

Learning Clean Generative Processes from Noisy Measurements

We study how to learn a generative model of clean images using only noisy observations. Diffusion processes and their generalizations to non-Gaussian noise require posterior predictions at every noise level. We first show that it is possible to learn these predictions from independent noisy views using Noise2Noise regression, but only at or above the noise level of the training data. To learn the predictions below that level, needed to generate clean images, we impose consistency: a prediction should equal its expectation after continuing the sampling process. We derive conditions under which consistency uniquely determines the continuation to the clean image distribution and verify them for Poisson, binomial, Gamma, and negative-binomial measurements. Experiments on CIFAR-10, photon-counting microscopy, and synthetic aperture radar show that the required posterior predictions can be learned from noisy views alone and that the consistency relation repairs the continuation beyond the noise level of the data, which severely deteriorates without consistency.

en

arXiv AI Papershace 14 días

There Is No Free Loop: When Does Looping a Trained Model Pay for Its Inference Compute?

Looping re-applies a block of layers so that each token passes through the block more than once. Adding a loop to a trained model is attractive because it leaves the parameter count, and hence the disk size, the memory footprint and the training cost unchanged. Unfortunately, every generated token pays for the loop. Re-applying m of N layers makes each token about (N+m)/N times as expensive. Loops that a model learns during large-scale pretraining justify this price, but whether loops added after pretraining justify this cost remains unclear. We compare a looped model with two simple baselines. We first compare it with the same model trained identically without the loop (equal parameters). Secondly, we compare it with the best unlooped use of the same inference floating-point operations (equal FLOPs). Across Qwen3 and Qwen3.5 models from 0.8B to 9B parameters, only the looped 0.8B model beats both baselines. At 4B and 9B, the looped model does not beat the first baseline, i.e. the same model trained without the loop. This is not because the benchmarks are too easy, since on the same benchmarks the unlooped 9B model still scores up to 13 points higher than the 4B model, showing that there is ample room for improvement. Introducing loops in an already pretrained model therefore only pays for small models with a fixed parameter budget.

en

arXiv AI Papershace 14 días

Empirical Evaluation of Quantum Kernel Methods for Classification: A Comparative Study with Classical Kernels

Quantum kernel methods are the most concrete near-term proposal for quantum machine learning on classical data, yet published benchmarks disagree on whether they outperform tuned classical kernels, and the disagreement cannot be resolved from the literature because the comparisons are not protocol-controlled. We argue that the obstacle is one of measurement, not theory: the usual setup fixes the quantum feature map's data scaling while grid-searching the classical kernel's hyperparameters, recomputes nothing per fold, and reports cost in simulator time. We therefore treat the comparison itself as the object of study and impose three controls. The bandwidth is tuned on the same grid budget as the classical γ, the kernel is recomputed inside every fold from fold-fitted preprocessing, and cost is counted in hardware circuit evaluations. On top of these we run a factorial design of 324 cells in which dimensionality, training size, class separability and data structure are explicit covariates, together with four standard datasets, a shots-versus-gate-noise ablation, and a cost sweep. The dominant effect we measure is not quantum: tuning the feature map's bandwidth is worth 0.120 accuracy on synthetic data and 0.339 on real data, exceeding every quantum-classical difference we observe. Tuned quantum kernels are then statistically indistinguishable from a tuned RBF on all four real datasets. The advantage on synthetic data is confined almost entirely to data generated by the feature map itself, and it shrinks with dimensionality (-0.0045 per qubit) while growing with sample size. Finite sampling costs an order of magnitude more accuracy than gate and readout noise, and a fidelity kernel needs 510^{7} circuits at n = 10,000. Protocol variance is thus sufficient to explain the field's contradictory conclusions.

en

arXiv AI Papershace 14 días

CausalXRL: Explainable Reinforcement Learning through Causal Graph Reasoning

Reinforcement learning is a powerful paradigm for training autonomous agents and has achieved impressive performance in complex environments. However, this success often comes at the cost of interpretability, diminishing trust and complicating efforts to debug and improve agent behavior. To address these challenges, we introduce CausalXRL, a novel framework for explainable reinforcement learning (XRL). A key feature of CausalXRL is its use of causal graph reasoning, which provides transparent, structured, multi-level explanations of agent decision-making. We validate CausalXRL through comprehensive case studies and a two-part evaluation: (1) a quantitative analysis of explanation fidelity and causal-structure learning efficiency in benchmark RL environments, and (2) a qualitative expert study assessing explainability in the real-time strategy (RTS) benchmark MicroRTS. The quantitative results show that CausalXRL can provide faithful explanations while efficiently learning causal structures, and the qualitative expert study suggests that participants found CausalXRL useful for inspecting high-level RTS strategies.

en

arXiv AI Papershace 14 días

A Neural Network Approach towards Real-Time Knot Point Density Estimates for the Nonlinear Quadrotor

Spline representations provide a compact way to approximate continuous trajectories using a finite number of parameters. When approximating trajectories generated by dynamical systems, the efficiency of a spline representation depends strongly on the placement of knot points. Uniform knot distributions often lead to inefficient representations, requiring either unnecessary control points or producing poor approximations in regions where the trajectory varies rapidly. In this work we propose a learning-based method for adaptive knot placement in B-spline approximations of trajectories generated by ordinary differential equations. A neural network learns a mapping from the initial condition of the system to quantities that characterize the local variation of the resulting trajectory. These estimates are used to construct nonuniform knot distributions that approximately equidistribute a local error indicator, reducing the number of control points needed to meet a prescribed accuracy.

en

arXiv AI Papershace 14 días

GhostPrefix: Synchronization-Free Universal Prefix Attacks for ASR Manipulation

We present GhostPrefix, the first synchronization-free universal adversarial prefix attack against automatic speech recognition (ASR) systems. Unlike prior audio attacks, which are typically crafted per utterance, require synchronization with the victim waveform, or target narrow objectives, GhostPrefix learns a short, reusable audio segment prepended to arbitrary live speech to exert real-time control over ASR outputs without knowledge of the future utterance or temporal alignment. A key challenge is maintaining stable adversarial effects across unknown, evolving speech contexts. GhostPrefix combines token-level losses with a novel representation-level objective for stable, input-agnostic effects. GhostPrefix supports three attack modes: (i) prompted prefix attacks for semantic steering, (ii) targeted prefix attacks for transcript hijacking, and (iii) untargeted prefix attacks for transcription degradation. To remain perceptually inconspicuous, we regularize prefixes toward benign environmental-sound distributions and improve robustness against over-the-air distortions. We evaluate GhostPrefix across 13 open-source ASR models from four families. In digital settings, it achieves average attack success rates of 97.6% for prompted steering and 93.4% for targeted hijacking, while driving word error rate to 1.18 under untargeted attacks. Over the air, it achieves success rates of 91.5% and 89.5% for prompted and targeted modes, respectively. These manipulations propagate into downstream AI-assisted pipelines, including voice assistants and AI note-taking tools, causing systematic semantic misinterpretations. GhostPrefix exposes synchronization-free universal prefixes as a practical, under-explored threat to real-world ASR and downstream applications. Artifacts and demos are available at https://ghostprefix.github.io/.

en

arXiv AI Papershace 14 días

Know When the Equation Is Wrong: Structural Uncertainty for LLM-Guided Symbolic Regression under Extrapolation

Large language models (LLMs) can help discover mathematical equations from data, but most LLM-guided symbolic regression methods focus on a single best equation without quantifying uncertainty over alternative functional forms. This is challenging because distinct equations can fit the observed data similarly while producing substantially different extrapolations. Our experiments on the LLM-SRBench dataset show that standard conformal intervals achieve nominal in-distribution coverage but frequently fail under extrapolation. To tackle this challenge, we introduce LLM-SR-UQ, a framework that retains competing equation structures and uses their disagreement to characterize uncertainty about the underlying functional form. We extend the standard fixed-width conformal approach to quantify structural intervals, which measure disagreement among candidate equations and identify where prediction intervals should widen. Extensive experiments show our framework substantially improves coverage attainment by incorporating structural disagreement while preserving in-distribution coverage. Our approach also shows good generalizability when it is applied to other symbolic-regression methods.

en

arXiv AI Papershace 14 días

Asynchronous Flow Map Language Models

The order in which tokens are predicted in language modeling has a profound impact on the quality of generated sequences. Autoregressive and masked diffusion models iteratively fill in tokens and can exploit such orderings across generation steps, but how to leverage an ordering for one-step parallel language modeling is unclear. We introduce asynchronous flow map language models (aFMLMs), which incorporate prediction order into one-step modeling by learning the flow maps of asynchronous flows. These flows specify an ordering through a tokenwise time schedule that controls the relative denoising progress of tokens. We formulate a general class of asynchronous flows that admit both prescribed and learned, state-adaptive schedules, show that they define valid flow maps, and develop new methods for their flow map distillation. The resulting models can generate all tokens in a single forward pass by internalizing the ordering encoded in the flow into their hidden computations. On Sudoku solving and code generation for GSM8K math word problems, aFMLM establishes new state-of-the-art one-step performance.

en

arXiv AI Papershace 14 días

When Does Jacobian Fine-Tuning Stabilize Neural PDE Emulators?

Neural emulators of chaotic PDEs are trained on one-step prediction but deployed autoregressively, where their errors compound. Jacobian-based regularisers aim to curb this growth, but in a chaotic system part of the growth is physical and cannot be trained away. We separate the two sources of this error with a first-order decomposition. We compare the emulator's error with a counterfactual that propagates the same one-step mistakes through the solver's Jacobian, giving an amplification ratio: the factor by which the model's own dynamics amplify its errors. We study 21 configurations across four PDEs, five architectures and two model sizes (8-10k and 33k parameters), with errors measured at one Lyapunov time. Jacobian fine-tuning reduced the error by more than 10% for 1 of 34 seeds whose amplification ratio was at most 1.25, and for all 20 seeds whose ratio was at least four; for the latter, the error fell by a factor of 8 (small models) and 55 (33k models) in geometric mean over seeds. The gain approximately equals the one-step accuracy cost of fine-tuning divided by the amplification ratio. Matching the solver's Jacobian along the model's own error direction (EDJM) outperformed random directions at the same cost on 28 of 35 seeds; input-noise injection performed comparably to EDJM.

en

arXiv AI Papershace 14 días

TurnTurk-GPT: Text-Based Turn-Taking Prediction for Turkish

This paper presents TurnTurk-GPT, a text-based turn-taking predictor for Turkish obtained by fine-tuning a Turkish GPT-2 model with the TurnGPT objective on synthetic and real dialogue, and evaluates it on real Turkish conversation against rule baselines, a zero-shot Turkish language model and a widely used multilingual end-of-turn detector. Evaluation uses word-level decision points on unpunctuated, lower-cased text approximating the output of streaming speech recognisers, recording-disjoint and speaker-disjoint folds, nested threshold selection and paired cluster bootstrap tests. TurnTurk-GPT outperforms the multilingual detector under both recording-disjoint and speaker-disjoint evaluation. It outperforms the zero-shot language model only when the test speakers occur in training; on unseen speakers the two are statistically indistinguishable. Removing the suffixes of the turn-final word reduces AP by up to 44%, against 2% for a sentence-internal word, locating the Turkish end-of-turn cue in verb-final morphology. Synthetic dialogue alone yields the weakest trained model, and when a speaker pauses the zero-shot and multilingual baselines remain close to chance. The model and the evaluation protocol are publicly released.

en

arXiv AI Papershace 14 días

TradeWar: A Benchmark for Execution-Grounded Trade Negotiation with Adaptive LLM Agents

Multi-turn negotiation remains a challenging problem for large language model (LLM) agents. While LLMs can conduct natural-language dialogue, existing benchmarks predominantly assess only agreement rates, overlooking whether negotiated commitments actually create value once executed. To address this gap, we introduce TradeWar, a multi-country simulation benchmark that translates negotiation dialogues into concrete tariff and trade-share policies, enforces capacity and network constraints during settlement, and evaluates outcomes via an Economic Health Index (EHI). To solve this execution-grounded task, we propose TACTIC (Trade-aware Adaptive Consolidation and Theory-of-mind for International Commerce), an agent that iteratively refines its negotiation strategy across episodes-without updating base LLM weights-by leveraging prior experience, opponent behavior, and past settlement outcomes. Despite using zero-shot chain-of-thought (CoT) prompting, PromptLLM attains a 75% agreement rate but only 57.1 final EHI (*100), whereas TACTIC reaches 68.8. Across 20 sequential episodes and three seeds, TACTIC's final EHI improves from 58.2 to 73.2, compared with 65.2 for an ablation without cross-episode updates. Under technology-controls and energy-crisis shocks, TACTIC also exhibits smaller performance degradation than PromptLLM and MemoryLLM. These results indicate that CoT deliberation and interaction memory alone do not fully address delayed execution feedback, whereas settlement-grounded adaptation improves long-horizon negotiation in TradeWar.

en

arXiv AI Papershace 14 días

AutoAnalysis: An automated, self-updating meta-analysis of human-AI collaboration

As generative artificial intelligence (AI) becomes more capable and widely used, understanding when and how humans should work with these systems is becoming increasingly important. However, the empirical evidence in this space offers conflicting recommendations, which leaves individuals and organizations without clear guidance on when to combine human and AI capabilities. Existing meta-analyses can help reconcile these findings, but they require substantial manual effort and cannot readily incorporate new evidence as it emerges. To address this challenge, we introduce AutoAnalysis: an automated, self-updating meta-analysis of when humans and AI work better together, and when they work better alone. AutoAnalysis' most recent run synthesizes 1,955 effect sizes from 420 papers. Across this body of work, we found that, on average, human-AI combinations performed significantly better than humans alone, but they performed significantly worse than the better of humans or AI alone. These findings highlight the important distinction between improving human performance and achieving human-AI synergy and underscore the need to evaluate collaboration against both human and AI performance alone. As new studies emerge, AutoAnalysis will integrate their findings into the meta-analysis, providing cumulative evidence about when humans and AI perform better together and how human-AI performance varies across AI agents, tasks, and processes.

en

arXiv AI Papershace 14 días

Less Loss, Fixed Weights: Logit Corrections for Forgetting

When a language model is fine-tuned to follow instructions, or further pretrained on a new subject, it often gets worse at predicting standard web text. This is known as "forgetting." We ask whether this loss can be undone by simple, fixed adjustments to the model's final token predictions. Across 14 pairs of base and derived models, from 1B to 7B parameters, we fitted a fixed bonus or penalty for each token (a vocabulary bias), alone or together with a temperature. On text held out from correction fitting, the largest top-4K bias recovery estimate is under 10%, with loss conditional on the retained vocabulary. Combining the bias with a temperature recovers up to 36%. Models further pretrained on math lose the most and recover almost nothing. Finally, in a controlled simulation where a model is damaged at random, with no planted bias or temperature change, the bias-plus-temperature fix still recovers 29%. Most of the real-model loss increase stays out of reach of our restricted fits.

en

arXiv AI Papershace 14 días

SecPO: Principled Adversarial Training for Prompt Injection Security

Recognized as a top threat to AI agents, prompt injection allows instructions from untrusted data (email, documents, websites) to manipulate the agent's operation. The existing best-defended LLM, Meta-SecAlign, is trained with the DPO loss but still suffers from attack success rates (ASRs) approaching 100% under adaptive prompt injections. One reason is that DPO anchors the model around the undefended reference model, which is insecure. We show an idealized reference model---an injection-invariant, utility-preserving oracle---can nevertheless be approximated, as its output distribution in training is interestingly accessible. This new observation motivates a theoretically principled SecPO loss. Another Meta-SecAlign shortcoming is that it trains on weak simulated attacks. We implement an efficient adversarial training (AT) pipeline that synthesizes strong injected inputs (best@240), and train the model on them. Our AT reduces ASRs by 1.5-14.5across three SoTA adaptive attacks, compared with Meta-SecAlign; the SecPO loss reduces them by a further 2.4-4.6compared with using DPO loss. In agentic workflows, our SecPO + AT Qwen3.6-27B is much more secure than Meta-SecAlign: 8.8% vs. 78.8% AgentDojo ASR@800 under the AttackerMovesSecond attack. Moreover, our defended model shows no noticeable reduction in benign utility in nine benchmarks covering agentic tool calling, instruction following, science, and math. We release our {//github.com/Sizhe-Chen/SecPO}{code}, ${//huggingface.co/datasets/Sizhe-Chen/Qwen3.6-27B-Instruct-SecPO-trainset}{training data} (with optimized injections from AT), and {//huggingface.co/Sizhe-Chen/Qwen3.6-27B-Instruct-SecPO}{model}$.

en

arXiv AI Papershace 14 días

DAMI: Dual-Axis Memory Initialization for Enhanced Long-Context Retrieval in Gated Linear Attention

Linear attention scales linearly with sequence length, but its finite-state memory can limit long-context retrieval. In gated linear recurrences, memory retention depends on state decay, which is shaped jointly by gate parameters A_log and dt_bias and by input activations. In controlled Gated DeltaNet (GDN) experiments, both parameters remain highly correlated with their initial values after training, highlighting the importance of initialization. Therefore, we introduce Dual-Axis Memory Initialization (DAMI), which jointly configures these parameters before training. DAMI assigns geometric timescales across heads for varied retrieval distances and a low-high-low input-sensitivity profile across layers. A small activation sample calibrates the GDN-family implementation to match both behavioral targets without adding trainable parameters. Across GDN, KDA, GDN2, Mamba-2, and RWKV-7, DAMI improves average RULER-style retrieval accuracy by 5.29 to 16.71 percentage points over the corresponding random-initialization baselines. Performance on language modeling, commonsense reasoning, and 2K real-world retrieval remains broadly comparable to those baselines. Our code is available at https://speedai.com/research/dami.

en

arXiv AI Papershace 14 días

ID2Pattern: Pretraining Generalizable Identity Behavior Patterns for Transaction Risk Foundation Models

Large-scale behavioral pretraining offers a promising route to generalizable representations in online payment systems. However, high-cardinality identity fields such as card, device, and account IDs, though rich in behavioral signals, remain underexploited. Their continually growing vocabularies and limited semantic meaning make existing vocabulary-based and semantics-based encodings ill-suited. Their generalizable signal lies not in specific values, but in the identity behavior patterns they exhibit across a transaction sequence. Building on this insight, we propose ID2Pattern, a pretraining framework that brings these patterns into transaction foundation models. ID2Pattern first encodes identity fields with a Local Anonymous ID (LAID) tokenizer, which captures how each identity appears and recurs within the sequence, independent of its specific value. It then pretrains a multi-scale pattern codebook on unlabeled sequences, which abstracts LAID sequences into discrete behavior patterns for downstream prediction. Experiments on an industrial dataset with over 100M sequences and six public datasets show that ID2Pattern consistently outperforms strong baselines, improving industrial by 4.13% and public datasets by 8.83% on average. Notably, ID2Pattern remains equally effective even when all identity values are unseen during training, highlighting its potential to generalize to new users and markets.

en

arXiv AI Papershace 14 días

New Lower Bounds for Private PAC Learning

We study realizable PAC learning under approximate differential privacy, and give a framework for proving lower bounds on its sample complexity. We construct classes with identical VC and Littlestone dimensions but different private sample complexities at common privacy and accuracy parameters, showing that these dimensions do not determine the sample complexity, even for improper learners. To address this we introduce the Independent Class Packing (ICP) parameter, the largest total sample complexity of classes in {G} that can be packed independently into a class {H}, and prove that every (, δ)-DP, (α, β)-PAC learner for {H}, possibly improper, uses Ω(ICP({H}, {G})/(α)) examples. Using this framework we prove tight lower bounds for axis-aligned rectangles, union of intervals and linear subspaces and improved lower bounds for halfspaces and convex polygons. Finally we also show that for product classes, and more generally when the same classes pack into {H} and cover its domain so that target labels can be used directly, ICP characterizes private PAC sample complexity up to logarithmic factors.

en

arXiv AI Papershace 14 días

When Does Test-Time Search Help World Action Models? An Empirical Audit of Candidate Selection and Simulator Integrity

World-action models increasingly apply test-time scaling, spending extra inference computation on candidate selection and adaptive allocation to improve control. Such gains are credible only when candidates carry exploitable signal, nominal compute savings become real speedup, and evaluation stays reliable. We audit all three on Cosmos Policy and RoboCasa with a premise-first, pre-registered methodology. On clean data, a learned value selector is 1.66 pp worse than random; a hindsight oracle gains only 1.40 pp, with an episode-level interval spanning zero; and 82% of states show identical candidate outcomes, leaving little selection headroom. Token pruning cuts per-forward FLOPs by up to 47% but yields at most a measured 1.15end-to-end speedup on the deployed path (1.25on a decode-excluded path), short of a 1.5target, and our adaptive-allocation experiment is inconclusive under a power gate. We further identify and fix a silent rendering corruption whose rate rises sharply along compute-heavy planning paths while leaving the official direct-policy result unaffected, and release a detector, fix, and scripts. These findings show why test-time scaling should be audited as a chain of premises (decision signal, realized efficiency, evaluationintegrity) rather than by success rate alone. Conclusions are scoped to the tested configuration.

en

arXiv AI Papershace 14 días

Prune Gradually, Adapt Sparsely: Efficient Pruning of Large Language Models at Extreme Sparsity

Weight pruning can reduce the storage and inference cost of large language models (LLMs), but preserving quality at high sparsity levels remains challenging. Existing methods face a quality-cost trade-off in this regime. One-shot pruning is cheap but degrades sharply as sparsity increases, even when followed by post-pruning recovery. A growing line of work instead optimizes the mask directly from the dense model under the end-to-end objective; among these methods, [Lee et al.,2026], the recent state of the art at extreme sparsity, reaches high quality but at the cost of model-sized optimizer state and auxiliary variables. To address this gap, we propose ETAPS, which prunes gradually and adapts sparsely: it tightens a global sparsity constraint in progressive stages while training a sparse delta with a small, fixed budget. Our key observation is that gradual weight removal keeps each pruning step small enough for sparse adaptation to compensate, allowing the adapted model to provide end-to-end gradients that guide the next pruning decision without model-sized optimizer state. Across six models from 125M to 13B parameters, ETAPS achieves lower perplexity than ELSA at every evaluated sparsity from 50% to 90%, with perplexity reductions of up to 42%. ETAPS is also far cheaper: at 90% sparsity, it prunes LLaMA2-7B on a single H200, where ELSA uses four, with 5.0smaller total persistent state and 49% fewer GPU-hours. Together, our findings point to a promising direction for achieving high quality in demanding regimes such as extreme sparsity at substantially lower cost.

en