News & Research

arXiv:2606.02544v1 Announce Type: cross Abstract: Diffusion large language models (dLLMs) have recently emerged as a promising alternative to autoregressive (AR) LLMs, offering faster inference through parallel or blockwise decoding. However, their masked language modeling formulation remains incompatible with standard token-level speculative decoding, one of the most effective acceleration techniques for AR models. In AR decoding, the causal mask preserves temporally valid token-level contexts, enabling a target model to verify multiple drafted tokens in a single forward pass. In contrast, dLLMs rely on mask tokens and bidirectional attention, causing the effective context to change across denoising steps and preventing direct token-level speculative verification. To bridge this gap, we propose a simple but effective speculative decoding algorithm for diffusion language models, named SimSD, which mainly adopts a plug-and-play masking strategy that equips dLLMs with temporally valid to

arXiv:2606.02559v1 Announce Type: cross Abstract: Post-training compression of Large Language Models (LLMs) removes entire architectural components, either deleting them or replacing them with fitted modules. Existing replacement-based methods share two design constraints: full-layer granularity and contiguous selection. We argue that this is overly restrictive: in fact, redundancy in pretrained transformers is not confined to contiguous regions, nor does it evenly distribute between Attention and FeedForward outputs, implying that different strategies best approximate different submodule types and that removable components need not cluster within contiguous depth ranges. Based on this intuition, we introduce SubFit (Submodule-level Fitted residual replacement), which compresses LLMs at the submodule level: Attention and FeedForward submodules are selected non-contiguously, and each receives its own lightweight fitted residual bypass. SubFit operates post-training and requires only cal

arXiv:2606.02569v1 Announce Type: cross Abstract: Video is temporally redundant: adjacent frames usually share most objects, background, and layout. Yet existing video multimodal large language models (video MLLMs) usually encode each sampled frame as an independent RGB image, causing visual tokens to repeat content already present in earlier frames. This suggests a more direct video interface: send a full reference frame only when the scene cannot be predicted well from prior context, and otherwise transmit a compact description of inter-frame changes. We call this interface a \emph{predictive visual code}, and instantiate it for video MLLMs as \textbf{AdaCodec}. AdaCodec spends full visual tokens on a reference frame only when its conditional predictive cost is high; otherwise, it encodes inter-frame changes, including motion and prediction residuals, as compact P-tokens. Across all eleven benchmarks, AdaCodec improves over the Qwen3-VL-8B per-frame RGB baseline at a matched visual-t

arXiv:2606.02578v1 Announce Type: cross Abstract: Recent multimodal large language models have demonstrated strong reasoning ability, yet their reliability as automated evaluators remains limited by a critical weakness: when visual evidence conflicts with textual cues, MLLM judges tend to reward plausible narratives over perceptually correct answers. We identify and systematically analyze this phenomenon, which we term Perceptual Judgment Bias. Through controlled visual perturbations, existing multimodal judges frequently anchor on the response text instead of their own visual perception, leading to inconsistent and non-verifiable evaluations. To address this issue, we introduce the Perceptually Perturbed Judgment Dataset, which constructs minimally edited counterfactual responses that isolate perceptual errors and enable verifiable supervision. Building on this dataset, we develop a unified training framework that combines a structured GRPO-based reward with a batch-ranking objective,

arXiv:2407.15510v2 Announce Type: replace Abstract: Abstraction is key to human and artificial intelligence as it allows one to identify common structure in otherwise distinct objects or situations. Anti-unification (or generalization) is the branch of theoretical computer science and artificial intelligence that studies abstraction and has found applications in areas such as inductive logic programming, program synthesis, and analogy-making. To date, anti-unification has been studied almost exclusively from a syntactic perspective. In this paper, we initiate an algebraic (i.e.\ semantic) theory of anti-unification in the general setting of universal algebra, thereby extending anti-unification from term-based representations to arbitrary algebras and beyond equational theories. In particular, we introduce the notions of algebraic generalization ordering and minimally general generalization, establish basic structural properties, prove compatibility with homomorphisms and isomorphisms,

arXiv:2409.18624v4 Announce Type: replace Abstract: Unsupervised learning methods have a soft inspiration in cognition models. To this day, the most successful unsupervised learning methods revolve around clustering samples in a mathematical space. In this paper we propose a primitive-based, unsupervised learning approach for decision-making inspired by a novel cognition framework. This representation-centric approach models the input space constructively as a distributed hierarchical structure in an input-agnostic way. We compared our approach with both current state-of-the-art unsupervised learning classification, with current state-of-the-art small and incomplete datasets classification, and with current state-of-the-art cancer type classification. We show how our proposal outperforms previous state-of-the-art. We also evaluate some cognition-like properties of our proposal where it not only outperforms the compared algorithms (even supervised learning ones), but it also shows a dif

arXiv:2411.05196v3 Announce Type: replace Abstract: This study presents DhondtXAI as a SHAP-independent, D'Hondt-based attribution framework for tabular XAI. Instead of model-native feature importance or SHAP values, DhondtXAI computes background-interventional removal effects, separates positive and negative evidence, forms optional feature alliances, applies optional thresholds, allocates seats via the D'Hondt rule, and projects onto the local model-output difference. Completeness is preserved by construction, with the projection residual ratio reported as a diagnostic. The method is evaluated on synthetic additive and interaction tests, correlated-feature perturbations, operator and apportionment ablations, projection-mode comparisons, logit-scale checks, repeated split validation, paired deletion tests, and two healthcare datasets: Wisconsin Diagnostic Breast Cancer (CatBoost) and early-stage diabetes risk prediction (XGBoost). SHAP serves only as an external comparator with aligne

arXiv:2502.04512v4 Announce Type: replace Abstract: AI advancements have been significantly driven by a combination of foundation models and curiosity-driven learning aimed at increasing capability and adaptability. Within this landscape, open-endedness, where AI agents autonomously and indefinitely generate novel behaviors, representations, or solutions, has gained increasing interest. This has become relevant in the context of self-evolving agents and long-horizon discovery. This position paper argues that the defining properties of open-ended AI systems introduce a distinct and underexplored class of safety challenges, including loss of predictability, emergent misalignment, and difficulties in maintaining effective control as systems evolve beyond their initial design assumptions, that must be addressed preemptively. These challenges differ qualitatively from those associated with task-bounded or static models and are unlikely to be addressed by existing safety frameworks alone, wh

arXiv:2503.03137v3 Announce Type: replace Abstract: Constructive neural combinatorial optimization (NCO) offers a promising paradigm for solving vehicle routing problems (VRPs) by directly learning to construct approximate optimal solutions, thereby reducing reliance on expert knowledge for algorithm design. However, scaling these methods to handle large-scale instances remains challenging due to high computational complexity. While recent dynamic search space reduction (SSR) methods can improve inference efficiency through geometric distance-based pruning, they often struggle on complex instances with non-uniform distributions or when optimal solutions rely heavily on non-spatial constraints. To address this critical issue, we propose Learning to Reduce (L2R), which is the first learning-based dynamic SSR framework. L2R learns to adaptively prioritize nodes by extracting patterns from problem-specific features to prune the search space at each step, enabling efficient and scalable sol

arXiv:2503.11832v5 Announce Type: replace Abstract: Recent vision language models (VLMs) have made remarkable strides in generative modeling with multimodal inputs, particularly text and images. However, their susceptibility to generating harmful content when exposed to unsafe queries raises critical safety concerns. While current alignment strategies primarily rely on supervised safety fine-tuning with curated datasets, we identify a fundamental limitation we call the ''safety mirage'', where supervised fine-tuning inadvertently reinforces spurious correlations between superficial textual patterns and safety responses, rather than fostering deep, intrinsic mitigation of harm. We show that these spurious correlations leave fine-tuned VLMs vulnerable even to a simple one-word modification-based attack, where substituting a single word in text queries with a spurious correlation-inducing alternative can effectively bypass safeguards. Additionally, these correlations contribute to the ove

arXiv:2504.03635v4 Announce Type: replace Abstract: Reasoning is a core capability of language models (LMs), yet it remains unclear how much model capacity is necessary to support reasoning during pretraining. In this work, we study the minimal parameter budget required for implicit reasoning, defined as the ability to infer new facts from learned knowledge without explicit chain-of-thought supervision. To isolate this phenomenon, we pretrain LMs from scratch in a controlled synthetic environment that mimics the structure and distribution of real-world knowledge graphs, and evaluate their ability to complete missing edges via multi-hop inference. From both a theoretical and an empirical perspective, we identify a scaling law linking this optimal parameter budget to a graph search entropy measure. Across a wide range of model sizes, training steps, and graph complexities, we show that an optimally sized language model can reliably reason over approximately 0.008 bits of information per

arXiv:2505.12741v3 Announce Type: replace Abstract: Language models are increasingly used not only as standalone predictors but also as components in larger inference systems, from test-time scaling to multi-agent collaboration. We study language model networks, where pre-trained language models serve as reusable nodes and intelligence emerges from their topology, communication, and optimization. Existing systems mostly communicate through natural language: easy to deploy, but discrete, inefficient, and hard to optimize from end-task supervision. We propose LMNet, a dense and differentiable realization of this paradigm. LMNet uses stripped LLMs as vertex modules and trainable seq2seq modules as communication edges, enabling intermediate nodes to exchange dense vectors while preserving natural-language input and output at the system boundary. By bypassing intermediate embedding and de-embedding, LMNet enables efficient information transfer, end-to-end gradient optimization, and learned

arXiv:2505.13273v2 Announce Type: replace Abstract: Large text-to-image diffusion models rarely expose reliable signals of when a prompt is likely to produce a poorly aligned generation, especially when training data is undisclosed. We study whether expert disagreement inside pre-trained mixture-of-experts (MoE) diffusion models can serve as a reliable estimate for epistemic uncertainty. We introduce EMoE, a training-free method that separates expert-specific computation paths at an early MoE layer, uses the same initial noise across paths, and measures variance among their latent representations after the first denoising step. This provides an uncertainty-aware prompt signal before full image generation, without auxiliary networks or training diffusion ensembles. On COCO and CC3M, EMoE ranks prompts by text-image alignment quality metrics more consistently than diffusion-specific and router-based baselines. We further apply EMoE to multilingual prompts and find systematic language-dep

arXiv:2505.18492v5 Announce Type: replace Abstract: Mathematical competition problems fall into two broad types: theorem proving, which asks for a proof of a given statement, and answer construction, which requires constructing a property-satifying object with proofs. With recent advances in large language models (LLMs), formal theorem-proving techniques have made substantial progress on theorem-proving problems, yet formal answer construction remains less studied. This exposes a mismatch between current LLM model families: general LLMs are strong at informal conjecturing but are expensive and unreliable at formal proof generation, whereas prover LLMs are cheap and optimized for formal proofs but weak at mathematical reasoning for proposing candidate answers. Moreover, Lean proof checking alone does not enforce that a constructed witness is a canonical answer: circular or non-closed-form witnesses can eliminate the existential quantifier while failing to constitute an admissible contes

arXiv:2505.19489v2 Announce Type: replace Abstract: The Linux kernel is a critical system, serving as the foundation for numerous systems. Bugs in the Linux kernel can cause serious consequences, affecting billions of users. Fault localization (FL), which aims at identifying the buggy code elements in software, plays an essential role in software quality assurance. While recent LLM agents have achieved promising accuracy in FL on recent benchmarks like SWE-bench, it remains unclear how well these methods perform in the Linux kernel, where FL is much more challenging due to the large-scale code base, limited observability, and diverse impact factors. In this paper, we introduce LinuxFLBench, a FL benchmark constructed from real-world Linux kernel bugs. We conduct an empirical study to assess the performance of state-of-the-art LLM agents on the Linux kernel. Our initial results reveal that existing agents struggle with this task, achieving a best top-1 accuracy of only 41.6% at file lev