Agregador

arXiv:2603.18280v4 Announce Type: replace-cross Abstract: Current alignment evaluation mostly measures whether models encode dangerous concepts and whether they refuse harmful requests. Both miss the layer where alignment often operates: routing from concept detection to behavioral policy. We study political censorship in Chinese-origin language models as a natural experiment, using probes, surgical ablations, and behavioral tests across nine open-weight models from five labs. Three findings follow. First, probe accuracy alone is non-diagnostic: political probes, null controls, and permutation baselines can all reach 100%, so held-out category generalization is the informative test. Second, surgical ablation reveals lab-specific routing. Removing the political-sensitivity direction eliminates censorship and restores accurate factual output in most models tested, while one model confabulates because its architecture entangles factual knowledge with the censorship mechanism. Cross-model transfer fails, indicating that routing geometry is model- and lab-specific. Third, refusal is no longer the dominant censorship mechanism. Within one model family, hard refusal falls to zero while narrative steering rises to the maximum, making censorship invisible to refusal-only benchmarks. These results support a three-stage descriptive framework: detect, route, generate. Models often retain the relevant knowledge; alignment changes how that knowledge is expressed. Evaluations that audit only detection or refusal therefore miss the routing mechanism that most directly determines behavior.

arXiv:2603.17651v2 Announce Type: replace-cross Abstract: Generative inbetweening (GI) seeks to synthesize realistic intermediate frames between the first and last keyframes beyond mere interpolation. As sequences become sparser and motions larger, previous GI models struggle with inconsistent frames with unstable pacing and semantic misalignment. Since GI involves fixed endpoints and numerous plausible paths, this task requires additional guidance gained from the keyframes and text to specify the intended path. Thus, we give semantic and temporal guidance from the keyframes and text onto each intermediate frame through Keyframe-anchored Attention Bias. We also better enforce frame consistency with Rescaled Temporal RoPE, which allows self-attention to attend to keyframes more faithfully. TGI-Bench, the first benchmark specifically designed for text-conditioned GI evaluation, enables challenge-targeted evaluation to analyze GI models. Without additional training, our method achieves state-of-the-art frame consistency, semantic fidelity, and pace stability for both short and long sequences across diverse challenges.

arXiv:2603.17372v3 Announce Type: replace-cross Abstract: Large vision-language models (VLMs) often exhibit weakened safety alignment with the integration of the visual modality. Even when text prompts contain explicit harmful intent, adding an image can substantially increase jailbreak success rates. In this paper, we observe that VLMs can clearly distinguish benign inputs from harmful ones in their representation space. Moreover, even among harmful inputs, jailbreak samples form a distinct internal state that is separable from refusal samples. These observations suggest that jailbreaks do not arise from a failure to recognize harmful intent. Instead, the visual modality shifts representations toward a specific jailbreak state, thereby leading to a failure to trigger refusal. To quantify this transition, we identify a jailbreak direction and define the jailbreak-related shift as the component of the image-induced representation shift along this direction. Our analysis shows that the jailbreak-related shift reliably characterizes jailbreak behavior, providing a unified explanation for diverse jailbreak scenarios. Finally, we propose a defense method that enhances VLM safety by removing the jailbreak-related shift (JRS-Rem) at inference time. Experiments show that JRS-Rem provides strong defense across multiple scenarios while preserving performance on benign tasks. Code is available at https://github.com/LeeQueue513/JRS-Rem.

arXiv:2603.11413v5 Announce Type: replace-cross Abstract: A recent Nature Medicine study reported that ChatGPT Health under-triages 51.6% of emergencies and concluded that consumer-facing AI triage poses safety risks. Its protocol, however, was an exam-style scaffold (forced A/B/C/D output, knowledge suppression, no clarifying questions) unlike how consumers use health chatbots. We ask how much of the headline error rate is a property of the measurement rather than of the models. In a first, mechanistic study, five frontier LLMs on a 17-scenario bank scored 6.4 points higher under naturalistic patient-style messages than under the constrained scaffold (p=0.015). In a second, faithful replication we ran the authors' own 60 released vignettes through six frontier models under four matched formats, with clinician validation of the rewrites and a blinded clinician audit of the LLM adjudicators. Here the direction reversed: free-text rewrites scored slightly below the exact prompt (78.7% vs 81.8%, p=0.020), removing only the answer scaffold changed little (80.6% vs 81.4%, p=0.81), and naturalistic input with a forced categorical answer beat both the exact prompt (84.4%, p=0.023) and free text (p=0.0001). On the four vignettes defining the original emergency rate, under-triage was 17% with the exact scaffold, 50% in free text, and 21% with a forced letter on the same message; every free-text "under-triage" was a same-day recommendation scored C rather than D, and 60% made escalation conditional on information the patient was asked to check, behavior a single-turn benchmark cannot score. The headline rate did not reproduce as a stable cross-model finding; much of it depends on output format and on mapping prose onto a four-point scale. Each format elicits a separate response, so the design measures format sensitivity, not format versus capability. Safety claims should report sensitivity to input wording, output format and adjudication.

arXiv:2603.00889v2 Announce Type: replace-cross Abstract: Large Language Models (LLMs) have exhibited remarkable reasoning capabilities, largely enabled by supervised fine-tuning (SFT)- and reinforcement learning (RL)-based post-training on high-quality reasoning data. However, reproducing and extending these capabilities in open and scalable settings is hindered by three data-centric challenges: (1) the cold-start problem, arising from the lack of seed datasets with detailed, long Chain-of-Thought (CoT) trajectories needed to initialize reasoning policies; (2) limited domain coverage, as most existing open-source reasoning datasets are concentrated in mathematics, with limited coverage of broader scientific disciplines; and (3) the annotation bottleneck, where the difficulty of frontier-level reasoning tasks makes reliable human annotation prohibitively expensive or infeasible. To address these challenges, we introduce CHIMERA, a compact synthetic reasoning dataset of 9K samples designed to support generalizable reasoning across domains. CHIMERA is constructed with three key properties: (1) it provides rich, long CoT reasoning trajectories synthesized by state-of-the-art reasoning models; (2) it has broad and structured coverage, spanning 8 major scientific disciplines and over 1K fine-grained topics organized via a model-generated hierarchical taxonomy; and (3) it employs a fully automated, scalable evaluation pipeline that uses strong reasoning models to cross-validate both problem validity and answer correctness, removing the reliance on human annotations. We use CHIMERA to post-train a 4B Qwen3 model with SFT and RL. Despite the dataset's modest size, the resulting model achieves strong performance on a suite of challenging reasoning benchmarks, including GPQA-Diamond, AIME 24/25/26, HMMT 25 and Humanity's Last Exam, approaching or matching the performance of substantially larger models such as DeepSeek-R1 and Qwen3-235B.

arXiv:2601.22628v3 Announce Type: replace-cross Abstract: Test-Time Training offers a promising way to improve the reasoning ability of large language models (LLMs) by adapting the model using only the test questions. However, existing methods struggle with difficult reasoning problems for two reasons: raw test questions are often too difficult to yield high-quality pseudo-labels, and the limited size of test sets makes continuous online updates prone to instability. To address these limitations, we propose TTCS, a co-evolving test-time training framework. Specifically, TTCS initializes two policies from the same pretrained model: a question synthesizer and a reasoning solver. These policies evolve through iterative optimization: the synthesizer generates progressively challenging question variants conditioned on the test questions, creating a structured curriculum tailored to the solver's current capability, while the solver updates itself using self-consistency rewards computed from multiple sampled responses on both original test and synthetic questions. Crucially, the solver's feedback guides the synthesizer to generate questions aligned with the model's current capability, and the generated question variants in turn stabilize the solver's test-time training. Experiments show that TTCS consistently strengthens the reasoning ability on challenging mathematical benchmarks and transfers to general-domain tasks across different LLM backbones, highlighting a scalable path towards dynamically constructing test-time curricula for self-evolving. Our code and implementation details are available at https://github.com/XMUDeepLIT/TTCS.

arXiv:2601.21669v2 Announce Type: replace-cross Abstract: Many reinforcement learning (RL) problems admit multiple terminal solutions of comparable quality, where the goal is not to identify a single optimum but to represent a diverse set of high-quality outcomes. Nevertheless, policies trained by standard expected return maximization routinely collapse onto a small subset of outcomes, a phenomenon commonly attributed to insufficient exploration or weak regularization. We show that this explanation is incomplete: outcome level mode collapse is a structural consequence of the expected-return objective itself. Under idealized learning dynamics, the log-probability ratio between any two outcomes evolves linearly in their reward difference, implying exponential ratio divergence and inevitable collapse independent of the exploration strategy, entropy regularization, or optimization algorithm. We identify the source of this pathology as the probability multiplier inside the expectation and propose a minimal correction: inverse probability scaling, which removes outcome-frequency amplification from the learning signal, fundamentally changes the learning dynamics, and provably yields reward-proportional terminal distributions, preventing collapse in multimodal settings. We instantiate this principle in Group Relative Policy Optimization (GRPO) as a drop-in modification, IPS-GRPO, requiring no auxiliary models or architectural changes. Across different reasoning and molecular generation tasks, IPS-GRPO consistently reduces outcome-level mode collapse while matching or exceeding baseline performance, suggesting that correcting the objective rather than adding exploration heuristics is key to reliable multimodal policy optimization.

arXiv:2601.11663v2 Announce Type: replace-cross Abstract: Post-training quantization (PTQ) methods for large language models compress pretrained weights to low bit-widths without retraining by optimizing layer-local proxy objectives. Activation-aware scaling methods such as AWQ protect input channels with large activation norms, whereas second-order error-compensation methods such as GPTQ redistribute rounding error using the inverse of the unweighted input covariance matrix. We show that both paradigms -- along with diagonal Fisher quantization, block reconstruction, and Hadamard rotation methods -- are special cases of minimizing a single loss-derived quadratic form governed by gradient-weighted activation sensitivity. From a Gauss-Newton expansion of the task loss at a linear layer $Y = X W^\top$, we prove that under isotropic output curvature the expected excess loss from weight perturbation $\Delta W$ reduces to $\frac{1}{2 d_{\mathrm{out}}}\operatorname{Tr}(\Delta W H_D \Delta W^\top)$, where $H_D = \frac{1}{n} X^\top D X$ weights each calibration token's outer product by its squared output-gradient norm $D_{tt} = \|\nabla_{y_t}\mathcal{L}\|_2^2$. Within this framework, AWQ corresponds to a diagonal approximation of $H_D$ with uniform token weights ($D = I_n$) optimized via channel scaling, while GPTQ corresponds to Cholesky error compensation on $H_D$ under $D = I_n$. We derive closed-form optimal scaling exponents for diagonal sensitivity protection, formulate sensitivity-weighted Cholesky quantization (SAQ-GPTQ) on $H_D$, and prove bounds on token-gradient heterogeneity bias, cross-layer error amplification across residual blocks, and finite-sample calibration concentration. Controlled numerical simulations confirm that under token-gradient heterogeneity, replacing $X^\top X$ with $X^\top D X$ preserves $>99\%$ top-channel saliency recall and reduces INT3/INT4 excess loss by roughly $3\times$ over unweighted GPTQ.

arXiv:2601.09736v2 Announce Type: replace-cross Abstract: Linear Genetic Programming (LGP) is a powerful technique that allows for a variety of problems to be solved using a linear representation of programs. However, there still exists some limitations to the technique, such as the need for humans to explicitly map registers to actions. This thesis proposes a novel approach that uses Q-Learning on top of LGP, Reinforced Linear Genetic Programming (RLGP) to learn the optimal register-action assignments. In doing so, we introduce a new framework "linear-gp" written in memory-safe Rust that allows for extensive experimentation for future works.

arXiv:2512.13886v3 Announce Type: replace-cross Abstract: Post-training pruning cuts the inference cost of large language models, yet it faces a trade-off: simple heuristics that zero weights are fast but degrade accuracy, while principled joint optimization methods recover accuracy but are computationally infeasible at modern scale. One-shot methods such as SparseGPT sit in between and apply efficient but approximate weight updates. To close this gap, we introduce OPTIMA, a practical one-shot post-training pruning method that balances accuracy and scalability. OPTIMA casts layer-wise weight reconstruction after mask selection as independent, column-wise quadratic programs (QPs) that share one Hessian per weight matrix. Solving these QPs yields, at solver convergence, the per-column minimizer of the layer-wise reconstruction objective under the shared Hessian, and the shared Hessian makes the QPs easy to batch on accelerators. We implement an accelerator-friendly QP solver that accumulates one Hessian per weight matrix and solves many small QPs in parallel with a convergence certificate per column, enabling one-shot post-training pruning at scale on a single accelerator without fine-tuning. OPTIMA integrates with existing mask selectors and improves average zero-shot accuracy in 40 of 49 (model, mask selector, sparsity) settings across the LLaMA and Gemma families, with mean gains of +0.61, +0.56, and +1.01 points at 50% unstructured, 2:4, and 60% unstructured sparsity, and up to +3.97 points at 60% sparsity, while matching ADMM-based exact reconstruction on its own masks. On an NVIDIA H100, OPTIMA prunes an 8B-parameter transformer end-to-end in 40 hours and 60 GB of peak GPU memory. This shows that principled column-wise QP reconstruction is a practical alternative to heuristic weight updates in one-shot post-training pruning.

arXiv:2512.09496v2 Announce Type: replace-cross Abstract: Unequal representation of demographic groups in training data poses challenges to model generalisation across populations. Standard practice assumes that balancing subgroup representation always optimises performance. However, recent empirical results contradict this assumption: in some cases, imbalanced data distributions actually improve subgroup performance, and in others, subgroup performance is unaffected by the absence of an entire subgroup during training. We propose the latent separation hypothesis, which states that a fine-tuned model's dependence on subgroup representation is determined by the degree of separation between subgroups in the latent space of the pre-trained model. We formalise this hypothesis, provide theoretical analysis, validate it empirically across four vision and language datasets, and present preliminary evidence of a causal role through regularisation experiments. Finally, we show how quantitative analysis of latent subgroup separation can inform practical data collection and balancing decisions when fine-tuning foundation models.

arXiv:2511.18354v2 Announce Type: replace-cross Abstract: AI agents, not humans, are becoming the web's primary consumers, yet the web still ships content structured for human browsing, with no native way for publishers to expose the semantic units within it. Access is therefore mediated by a handful of AI providers that crawl, chunk, and index content unilaterally, leaving publishers little control over how their content is represented, retrieved, or attributed. We argue this is a structural problem, not one better chunking or retrieval algorithms can solve. The web lacks a native layer for publisher-controlled semantic discovery and retrieval. We propose AISTRAL, an AI-Native Internet architecture that supplies this layer. Publishers expose semantically meaningful units (e.g., passages, claims, tables) through explicit, policy-governed interfaces they control, and a web-native semantic resolver, analogous to DNS but operating over meaning rather than location, lets agents discover relevant sources before retrieving fine-grained content directly from publishers. Our end-to-end prototype matches or exceeds representative centralized retrieval pipelines in answer quality, with up to 4.2x higher accuracy on large, specialized document collections. Against a public web-search agent, it answers up to 4.2x more questions correctly, while moving 2.9-41x less data per query. Because retrieval terminates at publisher-operated servers, attribution, provenance, and usage policies travel with the content, realigning incentives across the web ecosystem.

arXiv:2511.16084v4 Announce Type: replace-cross Abstract: Hyperspectral image (HSI) classification typically involves large-scale data and computationally intensive training, which limits the practical deployment of deep learning models in real-world remote sensing tasks. This study introduces SpectralTrain, a universal, architecture-agnostic training framework that enhances learning efficiency by integrating curriculum learning (CL) with principal component analysis (PCA)-based spectral downsampling. By gradually introducing spectral complexity while preserving essential information, SpectralTrain enables efficient learning of spectral -- spatial patterns at significantly reduced computational costs. The framework is independent of specific architectures, optimizers, or loss functions and is compatible with both classical and state-of-the-art (SOTA) models. Extensive experiments on three benchmark datasets -- Indian Pines, Salinas-A, and the newly introduced CloudPatch-7 -- demonstrate strong generalization across spatial scales, spectral characteristics, and application domains. The results indicate consistent reductions in training time by 2-7x speedups with small-to-moderate accuracy deltas depending on backbone. Its application to cloud classification further reveals potential in climate-related remote sensing, emphasizing training strategy optimization as an effective complement to architectural design in HSI models. Code is available at https://github.com/mh-zhou/SpectralTrain.

arXiv:2511.01553v3 Announce Type: replace-cross Abstract: Edge AI must adapt to new classes and shifting data within a tight power budget, but online continual learning (learning from each sample before the next arrives) is where conventional processors are least efficient: each small update pays a full dispatch at full power. We present CLP-SNN, a spiking architecture that recasts the Continually Learning Prototypes (CLP) algorithm as event-driven neural dynamics: a self-normalizing local rule replaces CLP's global renormalization, prototype allocation becomes novelty-triggered neurogenesis, and correctness-modulated plasticity becomes reward-gated per-neuron metaplasticity. Deployed on Intel Loihi 2 as the first online continual learner on that chip, rehearsal-free CLP-SNN matches Replay and SLDA on OpenLORIS at 1-shot; at 25-shot, it stays within 0.8 percentage points of Replay and 6.2 points below covariance-based SLDA. Each prediction-and-update step on extracted features completes on-chip in 0.33 ms at 0.05 mJ, 7.3x lower latency and 341x lower energy than Replay on a Jetson Orin Nano.

arXiv:2510.06919v2 Announce Type: replace-cross Abstract: We present a method that models the evolution of an unbounded number of time series clusters by switching among an unknown number of regimes with linear dynamics. We develop a Bayesian non-parametric approach using a hierarchical Dirichlet process as a prior on the parameters of a Switching Linear Dynamical System and a Gaussian process prior to model the statistical variations in amplitude and temporal alignment within each cluster. By modeling the evolution of time series patterns, the method avoids unnecessary proliferation of clusters in a principled manner. We perform inference by formulating a variational lower bound for off-line and on-line scenarios, enabling efficient learning through optimization. We illustrate the versatility and effectiveness of the approach through several case studies of electrocardiogram analysis using publicly available databases.

arXiv:2510.03223v3 Announce Type: replace-cross Abstract: Large Language Models (LLMs) tend to generate a long reasoning chain when solving complex tasks. However, as the reasoning chain extends, critical intermediate steps and the original prompt will be buried in the context, receiving insufficient attention and leading to errors. In this work, we present ATAR, a novel reasoning method that leverages the inherent reasoning structure to steer LLM attention. Our experiments show that ATAR outperforms SOTA methods across six benchmarks, achieving up to 15.39% absolute improvement. Furthermore, with ATAR, "non-reasoning" models achieve comparable or even better performance compared to reasoning models of the same size in most benchmarks. Finally, our ablation studies show that the attention alignment component contributes significantly, and that these improvements are persist under different attentionsteering backends.

arXiv:2508.20976v2 Announce Type: replace-cross Abstract: Large audio-language models (LALMs) extend language understanding into the auditory domain, yet their ability to perform low-level listening, such as pitch and duration detection, remains underexplored. However, low-level listening is critical for real-world, out-of-distribution tasks where models must reason about unfamiliar sounds based on fine-grained acoustic cues. To address this gap, we introduce the World-of-Whale benchmark (WoW-Bench) to evaluate low-level auditory perception and cognition using marine mammal vocalizations. We use marine mammal vocalizations as out-of-distribution sound events to better assess models' low-level listening and so that the models do not rely on prior knowledge of the sound events. WoW-bench is composed of a Perception benchmark for categorizing novel sounds and a Cognition benchmark, inspired by Bloom's taxonomy, to assess the abilities to remember, understand, apply, and analyze sound events. For the Cognition benchmark, we additionally introduce distractor questions to evaluate whether models are truly solving problems through listening rather than relying on other heuristics. Experiments with state-of-the-art LALMs show performance far below human levels, indicating a need for stronger auditory grounding in LALMs.

arXiv:2506.10016v4 Announce Type: replace-cross Abstract: Multimodal generative models (MGMs) have rapidly expanded beyond text generation to encompass images, music, video, human motion, and 3D objects. This survey categorises six primary text-conditioned generative modalities and examines how four techniques, namely self-supervised learning, mixture of experts, reinforcement learning from human feedback, and chain-of-thought prompting, are adapted across them, distinguishing shared mechanisms from experimentally supported benefits. We organise evaluation around faithfulness, compositionality, and robustness, and interpret evidence within each study's protocol. Selected within-study comparisons reported in the reviewed literature support four findings: preference gains need not improve text relevance, and reward optimisation can reduce within-prompt diversity; intermediate plans require tests of their realisation in final outputs; expert granularity changes the mechanism and its storage, computation, and runtime costs; and gains against an optimisation score need confirmation outside that scoring path. We further analyse trustworthiness risks, including bias, privacy leakage, and misuse of generated media, and examine the evidence for mitigation strategies.

arXiv:2504.08596v4 Announce Type: replace-cross Abstract: Hallucination, the generation of non factual content by AI systems, poses serious risks in medical contexts, where errors can directly affect patient outcomes. We present MedHal, a large-scale dataset specifically designed to assess capabilities and train models on the task of hallucination detection in medical texts. Current hallucination detection methods face significant limitations when applied to specialized domains like medicine, where they can have disastrous consequences. MedHal addresses this issue by incorporating diverse medical text sources and tasks covering both intrinsic and extrinsic hallucinations, and by providing a substantial volume of data samples suitable for training medical hallucination detection models. We demonstrate MedHal's utility by training and evaluating a baseline medical hallucination detection model, showing improvements over general-purpose hallucination detection approaches. This resource enables more efficient evaluation and training of medical text generation systems while reducing reliance on costly expert review, potentially accelerating the development of medical AI research.

arXiv:2503.18462v2 Announce Type: replace-cross Abstract: Deep generative models (DGMs) have caused a paradigm shift in the field of machine learning, yielding noteworthy advancements in domains such as image synthesis, natural language processing, and other related areas. However, a comprehensive evaluation of these models that effectively captures the interplay among fidelity, diversity, and novelty in generated samples remains a persistent challenge. A recently proposed approach, the Feature Likelihood Divergence (FLD), offers a theoretically grounded and practical tool in this regard but faces notable computational limitations. In this paper, we propose PALATE, a novel framework designed to enhance the evaluation of DGMs by addressing the efficiency constraints of FLD. Our approach is based on a peculiar application of the law of total expectation to random variables representing accessible data. When integrated with the MMD baseline metric and the recently introduced DINOv3 feature extractor, PALATE provides a comprehensive evaluation framework that matches or exceeds state-of-the-art methods while offering superior computational efficiency and scalability to large-scale datasets. Through a series of experiments, we demonstrate that the PALATE enhancement constitutes a practical and computationally efficient framework for holistic DGM assessment, particularly in detecting sample memorization and evaluating generalization capabilities. The source code is provided as supplementary material.