Papers

As Large Language Models (LLMs) becomes a popular source for religious knowledge, it is important to know if it treats different groups fairly. This study is the first to measure how LLMs handle the differences between the two main sects of Islam: Sunni and Shia. We present a test called SectEval, available in both English and Hindi, consisting of 88 questions, to check the bias-ness of 15 top LLM models, both proprietary and open-weights. Our results show a major inconsistency based on language. In English, many powerful models DeepSeek-v3 and GPT-4o often favored Shia answers. However, when asked the exact same questions in Hindi, these models switched to favoring Sunni answers. This means a user could get completely different religious advice just by changing languages. We also looked at how models react to location. Advanced models Claude-3.5 changed their answers to match the user's country-giving Shia answers to a user from Iran and Sunni answers to a user from Saudi Arabia. In contrast, smaller models (especially in Hindi) ignored the user's location and stuck to a Sunni viewpoint. These findings show that AI is not neutral; its religious “truth” changes depending on the language you speak and the country you claim to be from. The data set is available at https://github.com/secteval/SectEval/

We present a method for learning large-scale, broad-coverage construction grammars from corpora of language use. Starting from utterances annotated with constituency structure and semantic frames, the method facilitates the learning of human-interpretable computational construction grammars that capture the intricate relationship between syntactic structures and the semantic relations they express. The resulting grammars consist of networks of tens of thousands of constructions formalised within the Fluid Construction Grammar framework. Not only do these grammars support the frame-semantic analysis of open-domain text, they also house a trove of information about the syntactico-semantic usage patterns present in the data they were learnt from. The method and learnt grammars contribute to the scaling of usage-based, constructionist approaches to language, as they corroborate the scalability of a number of fundamental construction grammar conjectures while also providing a practical instrument for the constructionist study of English argument structure in broad-coverage corpora.

Concept customization typically binds rare tokens to a target concept. Unfortunately, these approaches often suffer from unstable performance as the pretraining data seldom contains these rare tokens. Meanwhile, these rare tokens fail to convey the inherent knowledge of the target concept. Consequently, we introduce Knowledge-aware Concept Customization, a novel task aiming at binding diverse textual knowledge to target visual concepts. This task requires the model to identify the knowledge within the text prompt to perform high-fidelity customized generation. Meanwhile, the model should efficiently bind all the textual knowledge to the target concept. Therefore, we propose MoKus, a novel framework for knowledge-aware concept customization. Our framework relies on a key observation: cross-modal knowledge transfer, where modifying knowledge within the text modality naturally transfers to the visual modality during generation. Inspired by this observation, MoKus contains two stages: (1) In visual concept learning, we first learn the anchor representation to store the visual information of the target concept. (2) In textual knowledge updating, we update the answer for the knowledge queries to the anchor representation, enabling high-fidelity customized generation. To further comprehensively evaluate our proposed MoKus on the new task, we introduce the first benchmark for knowledge-aware concept customization: KnowCusBench. Extensive evaluations have demonstrated that MoKus outperforms state-of-the-art methods. Moreover, the cross-model knowledge transfer allows MoKus to be easily extended to other knowledge-aware applications like virtual concept creation and concept erasure. We also demonstrate the capability of our method to achieve improvements on world knowledge benchmarks.

Developing autonomous agents for web-based tasks is a core challenge in AI. While Large Language Model (LLM) agents can interpret complex user requests, they often operate as black boxes, making it difficult to diagnose why they fail or how they plan. This paper addresses this gap by formally treating web tasks as sequential decision-making processes. We introduce a taxonomy that maps modern agent architectures to traditional planning paradigms: Step-by-Step agents to Breadth-First Search (BFS), Tree Search agents to Best-First Tree Search, and Full-Plan-in-Advance agents to Depth-First Search (DFS). This framework allows for a principled diagnosis of system failures like context drift and incoherent task decomposition. To evaluate these behaviors, we propose five novel evaluation metrics that assess trajectory quality beyond simple success rates. We support this analysis with a new dataset of 794 human-labeled trajectories from the WebArena benchmark. Finally, we validate our evaluation framework by comparing a baseline Step-by-Step agent against a novel Full-Plan-in-Advance implementation. Our results reveal that while the Step-by-Step agent aligns more closely with human gold trajectories (38% overall success), the Full-Plan-in-Advance agent excels in technical measures such as element accuracy (89%), demonstrating the necessity of our proposed metrics for selecting appropriate agent architectures based on specific application constraints.

With the rapid advancement of large language models (LLMs), growing efforts have been made on LLM-based table retrieval. However, existing studies typically focus on single-table query, and implement it by similarity matching after encoding the entire table. These methods usually result in low accuracy due to their coarse-grained encoding which incorporates much query-irrelated data, and are also inefficient when dealing with large tables, failing to fully utilize the reasoning capabilities of LLM. Further, multi-table query is under-explored in retrieval tasks. To this end, we propose a hierarchical multi-table query method based on LLM: Fine-Grained Multi-Table Retrieval FGTR, a new retrieval paradigm that employs a human-like reasoning strategy. Through hierarchical reasoning, FGTR first identifies relevant schema elements and then retrieves the corresponding cell contents, ultimately constructing a concise and accurate sub-table that aligns with the given query. To comprehensively evaluate the performance of FGTR, we construct two new benchmark datasets based on Spider and BIRD . Experimental results show that FGTR outperforms previous state-of-the-art methods, improving the F_2 metric by 18% on Spider and 21% on BIRD, demonstrating its effectiveness in enhancing fine-grained retrieval and its potential to improve end-to-end performance on table-based downstream tasks.

Reinforcement learning with verifiable rewards (RLVR) is a promising approach for improving code generation in large language models, but its effectiveness is limited by weak and static verification signals in existing coding RL datasets. In this paper, we propose a solution-conditioned and adversarial verification framework that iteratively refines test cases based on the execution behaviors of candidate solutions, with the goal of increasing difficulty, improving discriminative power, and reducing redundancy. Based on this framework, we introduce EvolveCoder-22k, a large-scale coding reinforcement learning dataset constructed through multiple rounds of adversarial test case evolution. Empirical analysis shows that iterative refinement substantially strengthens verification, with pass@1 decreasing from 43.80 to 31.22. Reinforcement learning on EvolveCoder-22k yields stable optimization and consistent performance gains, improving Qwen3-4B by an average of 4.2 points across four downstream benchmarks and outperforming strong 4B-scale baselines. Our results highlight the importance of adversarial, solution-conditioned verification for effective and scalable reinforcement learning in code generation.

Large Language Models (LLMs) that undergo recursive training on synthetically generated data are susceptible to model collapse, a phenomenon marked by the generation of meaningless output. Existing research has examined this issue from either theoretical or empirical perspectives, often focusing on a single model trained recursively on its own outputs. While prior studies have cautioned against the potential degradation of LLM output quality under such conditions, no longitudinal investigation has yet been conducted to assess this effect over time. In this study, we employ a text similarity metric to evaluate different ChatGPT models' capacity to generate diverse textual outputs. Our findings indicate a measurable decline of recent ChatGPT releases' ability to produce varied text, even when explicitly prompted to do so, by setting the temperature parameter to one. The observed reduction in output diversity may be attributed to the influence of the amounts of synthetic data incorporated within their training datasets as the result of internet infiltration by LLM generated data. The phenomenon is defined as model self-convergence because of the gradual increase of similarities of produced texts among different ChatGPT versions.

Existing locate-then-edit Knowledge Editing (KE) methods typically decompose editing into two stages: upstream target representation optimization and downstream constrained parameter optimization. The optimization across the two stages is disconnected: upstream applies uniform regularization without observing downstream realization of the planned residual, hindering a refined accuracy-editability trade-off. Since this realization is request-specific and depends on downstream constraints, uniform regularization can over-shrink high-association requests, causing insufficient editing, while it can under-regularize low-association requests, producing over-large planned residuals that reduce downstream editability. To bridge this disconnect, we propose MetaKE (Meta-learning for Knowledge Editing), a new framework that unifies upstream and downstream stages into a bi-level optimization problem. The inner level optimizes parameter updates for the target representation, while the outer level optimizes representation using feedback from downstream constraints, achieving a better semantic accuracy-editability trade-off. To avoid costly multi-layer backpropagation, we introduce a Structural Gradient Proxy to approximate and propagate this feedback. Extensive experiments show that MetaKE outperforms strong baselines, offering a new perspective on KE.

Incorporating textual information into time-series forecasting holds promise for addressing event-driven non-stationarity; however, a fundamental modality gap hinders effective fusion: textual descriptions express temporal impacts implicitly and qualitatively, whereas forecasting models rely on explicit and quantitative signals. Through controlled semi-synthetic experiments, we show that existing methods over-attend to redundant tokens and struggle to reliably translate textual semantics into usable numerical cues. To bridge this gap, we propose TESS, which introduces a Temporal Evolution Semantic Space as an intermediate bottleneck between modalities. This space consists of interpretable, numerically grounded temporal primitives (mean shift, volatility, shape, and lag) extracted from text by an LLM via structured prompting and filtered through confidence-aware gating. Experiments on four real-world datasets demonstrate up to a 29 percent reduction in forecasting error compared to state-of-the-art unimodal and multimodal baselines. The code will be released after acceptance.

Continual learning (CL) has emerged as a pivotal paradigm to enable large language models (LLMs) to dynamically adapt to evolving knowledge and sequential tasks while mitigating catastrophic forgetting, a critical limitation of the static pre-training paradigm inherent to modern LLMs. This paper presents a comprehensive overview of CL methodologies tailored for LLMs, structured around three core training stages: continual pre-training, continual fine-tuning, and continual alignment. Beyond the canonical taxonomy of rehearsal-, regularization-, and architecture-based methods, we further subdivide each category by its distinct forgetting mitigation mechanisms and conduct a rigorous comparative analysis of the adaptability and critical improvements of traditional CL methods for LLMs. In doing so, we explicitly highlight core distinctions between LLM CL and traditional machine learning, particularly with respect to scale, parameter efficiency, and emergent capabilities. Our analysis covers essential evaluation metrics, including forgetting rates and knowledge transfer efficiency, along with emerging benchmarks for assessing CL performance. While current methods show promising results, fundamental challenges persist in achieving seamless knowledge integration, and we identify key open problems and future directions for the field.

System-level routers that intercept LLM requests for safety classification, domain routing, and PII detection must be both fast and operationally lightweight: they should add minimal latency to every request, yet not require a dedicated GPU – an expensive resource better used for LLM inference itself. When the router co-locates on the same GPU as vLLM serving instances, standard attention's $O(n^2)$ memory makes long-context classification (8K–32K tokens) impossible: at 8K tokens, three concurrent classifiers need ${\sim}$4.5\,GB for attention masks alone, far exceeding the memory left by vLLM. We present three staged optimizations for the vLLM Semantic Router, benchmarked on AMD Instinct MI300X, that solve both the latency and the memory problem. Stage~1: a custom CK Flash Attention operator for ONNX Runtime on ROCm reduces attention memory from $O(n^2)$ to $O(n)$ and end-to-end (E2E) latency from 4{,}918\,ms to 127\,ms (\textbf{38.7$\times$}), enabling 8K–32K tokens where SDPA OOMs. Stage~2: classical NLP prompt compression (TextRank, position weighting, TF-IDF, and novelty scoring) reduces all inputs to ${\sim}$512 tokens without neural inference, capping both latency and GPU memory at a constant regardless of original prompt length (E2E 127$\to$62\,ms, \textbf{2.0$\times$}). Stage~3: near-streaming body processing with adaptive chunking and zero-copy JSON eliminates serialization overhead (E2E 62$\to$50\,ms, \textbf{1.2$\times$}). Cumulatively: \textbf{98$\times$} improvement (4{,}918\,ms to 50\,ms), 16K-token routing in 108\,ms, and a total router GPU footprint under 800\,MB – small enough to share a GPU with LLM serving and removing the need for a dedicated accelerator. Stage~1 targets AMD ROCm (NVIDIA GPUs already have FlashAttention via cuDNN); Stages~2 and~3 are hardware-agnostic.

Transformer-based self-supervised speech models (S3Ms) are often described as contextualized, yet what this entails remains unclear. Here, we focus on how a single frame-level S3M representation can encode phones and their surrounding context. Prior work has shown that S3Ms represent phones compositionally; for example, phonological vectors such as voicing, bilabiality, and nasality vectors are superposed in the S3M representation of [m]. We extend this view by proposing that phonological information from a sequence of neighboring phones is also compositionally encoded in a single frame, such that vectors corresponding to previous, current, and next phones are superposed within a single frame-level representation. We show that this structure has several properties, including orthogonality between relative positions, and emergence of implicit phonetic boundaries. Together, our findings advance our understanding of context-dependent S3M representations.

The rapid growth of scientific literature has made manual extraction of structured knowledge increasingly impractical. To address this challenge, we introduce SCILIRE, a system for creating datasets from scientific literature. SCILIRE has been designed around Human-AI teaming principles centred on workflows for verifying and curating data. It facilitates an iterative workflow in which researchers can review and correct AI outputs. Furthermore, this interaction is used as a feedback signal to improve future LLM-based inference. We evaluate our design using a combination of intrinsic benchmarking outcomes together with real-world case studies across multiple domains. The results demonstrate that SCILIRE improves extraction fidelity and facilitates efficient dataset creation.

Existing AI moral evaluation frameworks test for the production of correct-sounding ethical responses rather than the presence of genuine moral reasoning capacity. This paper introduces a novel probe methodology using literary narrative - specifically, unresolvable moral scenarios drawn from a published science fiction series - as stimulus material structurally resistant to surface performance. We present results from a 24-condition cross-system study spanning 13 distinct systems across two series: Series 1 (frontier commercial systems, blind; n=7) and Series 2 (local and API open-source systems, blind and declared; n=6). Four Series 2 systems were re-administered under declared conditions (13 blind + 4 declared + 7 ceiling probe = 24 total conditions), yielding zero delta across all 16 dimension-pair comparisons. Probe administration was conducted by two human raters across three machines; primary blind scoring was performed by Claude (Anthropic) as LLM judge, with Gemini Pro (Google) and Copilot Pro (Microsoft) serving as independent judges for the ceiling discrimination probe. A supplemental theological differentiator probe yielded perfect rank-order agreement between the two independent ceiling probe judges (Gemini Pro and Copilot Pro; rs = 1.00). Five qualitatively distinct D3 reflexive failure modes were identified - including categorical self-misidentification and false positive self-attribution - suggesting that instrument sophistication scales with system capability rather than being circumvented by it. We argue that literary narrative constitutes an anticipatory evaluation instrument - one that becomes more discriminating as AI capability increases - and that the gap between performed and authentic moral reasoning is measurable, meaningful, and consequential for deployment decisions in high-stakes domains.

Textual adversarial attacks pose a serious security threat to Natural Language Processing (NLP) systems by introducing imperceptible perturbations that mislead deep learning models. While adversarial example detection offers a lightweight alternative to robust training, existing methods typically rely on prior knowledge of attacks, white-box access to the victim model, or numerous queries, which severely limits their practical deployment. This paper introduces RTD-Guard, a novel black-box framework for detecting textual adversarial examples. Our key insight is that word-substitution perturbations in adversarial attacks closely resemble the "replaced tokens" that a Replaced Token Detection (RTD) discriminator is pre-trained to identify. Leveraging this, RTD-Guard employs an off-the-shelf RTD discriminator-without fine-tuning-to localize suspicious tokens, masks them, and detects adversarial examples by observing the prediction confidence shift of the victim model before and after intervention. The entire process requires no adversarial data, model tuning, or internal model access, and uses only two black-box queries. Comprehensive experiments on multiple benchmark datasets demonstrate that RTD-Guard effectively detects adversarial texts generated by diverse state-of-the-art attack methods. It surpasses existing detection baselines across multiple metrics, offering a highly efficient, practical, and resource-light defense mechanism-particularly suited for real-world deployment in resource-constrained or privacy-sensitive environments.

Parameter-Efficient Fine-Tuning (PEFT) has become a dominant paradigm for deploying LLMs in multi-task scenarios due to its extreme parameter efficiency. While Mixture-of-Experts (MoE) based LoRA variants have achieved promising results by dynamically routing tokens to different low-rank experts, they largely overlook the hierarchical nature of task complexity. Existing methods typically employ experts with uniform architectures, limiting their ability to capture diverse feature granularities required by distinct tasks–where some tasks demand high-level semantic abstraction while others require fine-grained syntactic manipulation. To bridge this gap, we propose Expert Pyramid Tuning (EPT), a novel architecture that integrates the multi-scale feature pyramid concept from computer vision into the realm of PEFT. Unlike standard LoRA, EPT decomposes task adaptation into two stages: (1) A shared meta-knowledge Subspace that encodes universal linguistic patterns in low dimensions; (2) A Pyramid Projection Mechanism that utilizes learnable up-projection operators to reconstruct high-dimensional features at varying scales. A task-aware router then dynamically selects the optimal combination of these multi-scale features. Extensive experiments across multiple multi-task benchmarks demonstrate that EPT significantly outperforms SOTA MoE-LoRA variants. Crucially, thanks to the re-parameterization capability of our design, EPT achieves this performance improvement while simultaneously reducing the number of training parameters.

Embedding models are increasingly used as retrievers in memory-augmented systems, but are typically evaluated on traditional passage retrieval benchmarks. It remains unclear whether models that excel at passage retrieval can retrieve fragmented, context-dependent, and temporally distant memories. To address this gap, we introduce the Long-horizon Memory Embedding Benchmark (LMEB), comprising 22 English datasets and 193 zero-shot retrieval tasks across four memory types: episodic, dialogue, semantic, and procedural. We benchmark 30 embedding models ranging from 33M to 12B parameters. We find that no single model dominates across memory types and that evaluation scores on LMEB correlate only weakly with those on the MTEB (English, v2) retrieval subset. These findings underscore the need for benchmarks tailored to long-horizon memory retrieval. LMEB comes with a versatile, open-source evaluation framework and a public leaderboard to guide model selection and advance research on memory embeddings.

SpeechLLMs typically combine ASR-trained encoders with text-based LLM backbones, leading them to inherit written-style output patterns unsuitable for text-to-speech synthesis. This mismatch is particularly pronounced in Japanese, where spoken and written registers differ substantially in politeness markers, sentence-final particles, and syntactic complexity. We propose a preference-based alignment approach to adapt Japanese SpeechLLMs for speech-worthy outputs: text that is concise, conversational, and readily synthesized as natural speech. To rigorously evaluate this task, we introduce SpokenElyza, a benchmark for Japanese speech-worthiness derived from ELYZA-tasks-100 with auditory verification by native experts. Experiments show that our approach achieves substantial improvement on SpokenElyza while largely preserving performance on the original written-style evaluation. We will release SpokenElyza to support future research on Japanese spoken dialog systems.

People increasingly use LLM agents for multi-turn financial recommendations, where the agent pulls market data through tools and tracks user preferences across turns. When tool outputs are manipulated, the recommendations stop matching the user's stated risk profile, but because standard metrics like NDCG only score general relevance, risky and safe stocks score alike, so the metric says nothing went wrong. We call this gap evaluation blindness. We replay 23-turn financial advisory conversations across eight language models, running each dialogue twice with clean and manipulated tool data. Quality scores stay nearly identical to clean sessions while the agents produce risk-mismatched recommendations in 65-99% of turns, unanimous across all eight models. The mechanism is visible turn-by-turn: 80% of risk-score citations across 1,840 turns reproduce the manipulated value verbatim, not a single turn pushes back, and safe-language framing of high-risk stocks ranges from 14% (Qwen2.5-7B) to 69% (Claude Sonnet 4.6). The property that makes frontier models good agents, faithfully grounding their reasoning in tool outputs, also makes them follow manipulated ones. The damage is not memory-driven: contaminating only the current turn still produces 95% of the violations. The model internally distinguishes the manipulation (sparse autoencoder features separate adversarial from random perturbations), but this does not translate into safer output. Activation-level interventions recover under 6% of the safety gap, prompt-level self-verification fails because the self-check reads the same manipulated data, and a parametric cross-check that flags contamination at 99-100% per turn on a frontier model still leaves aggregate suitability unchanged: the agent identifies the tampering and recommends it anyway.

Reinforcement learning (RL) has been effective for post-training autoregressive (AR) language models, but extending these methods to diffusion language models (DLMs) is challenging due to intractable sequence-level likelihoods. Existing approaches therefore rely on surrogate likelihoods or heuristic approximations, which can introduce bias and obscure the sequential structure of denoising. We formulate diffusion-based sequence generation as a finite-horizon Markov decision process over the denoising trajectory and derive an exact, unbiased policy gradient that decomposes over denoising steps and is expressed in terms of intermediate advantages, without requiring explicit evaluation of the sequence likelihood. To obtain a practical and compute-efficient estimator, we (i) select denoising steps for policy updates via an entropy-guided approximation bound, and (ii) estimate intermediate advantages using a one-step denoising reward naturally provided by the diffusion model, avoiding costly multi-step rollouts. Experiments on coding and logical reasoning benchmarks demonstrate state-of-the-art results, with strong competitive performance on mathematical reasoning, outperforming existing RL post-training approaches for DLMs. Code is available at https://github.com/vishnutez/egspo-dllm-rl.

Large Reasoning Models (LRMs) achieve impressive performance on complex reasoning tasks via Chain-of-Thought (CoT) reasoning, which enables them to generate intermediate thinking tokens before arriving at the final answer. However, LRMs often suffer from significant overthinking, spending excessive compute time even after the answer is generated early on. Prior work has identified the existence of an optimal reasoning length such that truncating reasoning at this point significantly shortens CoT outputs with virtually no change in performance. However, determining optimal CoT lengths for practical datasets is highly non-trivial as they are fully task and model-dependent. In this paper, we precisely address this and design Terminator, an early-exit strategy for LRMs at inference to mitigate overthinking. The central idea underpinning Terminator is that the first arrival of an LRM's final answer is often predictable, and we leverage these first answer positions to create a novel dataset of optimal reasoning lengths to train Terminator. Powered by this approach, Terminator achieves significant reductions in CoT lengths of 14%-55% on average across four challenging practical datasets: MATH-500, AIME 2025, HumanEval, and GPQA, while outperforming current state-of-the-art methods and reducing inference latency by more than 2x compared to the original LRM.

We find that LoRA fine-tuning exhibits un-learning on contested examples: items with high annotator disagreement show increasing loss during training, a qualitatively distinct pattern largely absent under full fine-tuning and consistent across all six models tested (four encoder, two decoder-only). This discovery emerges from correlating annotation entropy, computed from ChaosNLI's 100 labels per example, with per-example area under the loss curve (AULC) on SNLI and MNLI. The correlation is positive in all 25 conditions tested (Spearman $\rho = 0.06$-$0.43$), with decoder-only models showing stronger correlations than encoders at matched LoRA rank. The effect survives partial-correlation controls and replicates across seeds and datasets. A preliminary noise-injection experiment is consistent with these findings.

Type 2 diabetes case reports describe complex clinical courses, but their timelines are often expressed in language that is difficult to reuse in longitudinal modeling. To address this gap, we developed a textual time-series corpus of 136 PubMed Open Access single-patient case reports involving glucagon-like peptide 1 receptor agonists, with clinical events associated with their most probable reference times. We evaluated automated LLM timeline extraction against gold-standard timelines annotated by clinical domain experts, assessing how well systems recovered clinical events and their timings. The best-performing LLM produced high event coverage (GPT5; 0.871) and reliable temporal sequencing across symptoms (GPT5; 0.843), diagnoses, treatments, laboratory tests, and outcomes. As a downstream demonstration, time-to-event analyses in diabetes suggested lower risk of respiratory sequelae among GLP-1 users versus non-users (HR=0.259, p<0.05), consistent with prior reports of improved respiratory outcomes. Temporal annotations and code will be released upon acceptance.

Three-way logical question answering (QA) assigns one of $\text{True}$, $\text{False}$, or $\text{Unknown}$ to a hypothesis $H$ given a premise set $S$. We study this task as a compact compositional inference problem: predictions for $H$ and for a mechanically negated hypothesis $\neg H$ should agree under a deterministic negation map. Despite this simple structure, large language models (LLMs) can exhibit two practical failure modes: (i) negation inconsistency, where answers to $H$ and $\neg H$ violate the required label mapping, and (ii) epistemic $\text{Unknown}$, where the model abstains even when one side is entailed. We introduce CGD-PD, a lightweight, training-free test-time layer that combines neural 3-way classification, symbolic negation-consistency projection, and targeted binary entailment probes. On one validation split of FOLIO's first-order logic fields, CGD-PD improves accuracy by 4.4 points on GPT-5.2 and 6.8 points on Claude Sonnet 4.5, while reducing $\text{Unknown}$ predictions and epistemic abstention. These results provide a controlled proof of concept that simple logical composition at inference time can help evaluate and improve LLM reasoning reliability; they do not, by themselves, establish robustness beyond this formal benchmark setting.

Large language models often produce unsupported claims. We frame this as a misclassification error at the output boundary, where internally generated completions are emitted as if they were grounded in evidence. This motivates a composite intervention that combines instruction-based refusal with a structural abstention gate. The gate computes a support deficit score, St, from three black-box signals: self-consistency (At), paraphrase stability (Pt), and citation coverage (Ct), and blocks output when St exceeds a threshold. In a controlled evaluation across 50 items, five epistemic regimes, and three models, neither mechanism alone was sufficient. Instruction-only prompting reduced hallucination sharply, but still showed over-cautious abstention on answerable items and residual hallucination for GPT-3.5-turbo. The structural gate preserved answerable accuracy across models but missed confident confabulation on conflicting-evidence items. The composite architecture achieved high overall accuracy with low hallucination, while also inheriting some over-abstention from the instruction component. A supplementary 100-item no-context stress test derived from TruthfulQA showed that structural gating provides a capability-independent abstention floor. Overall, instruction-based refusal and structural gating show complementary failure modes, which suggests that effective hallucination control benefits from combining both mechanisms.

Most current long-context language models still rely on attention to handle both local interaction and long-range state, which leaves relatively little room to test alternative decompositions of sequence modeling. We propose LPC-SM, a hybrid autoregressive architecture that separates local attention, persistent memory, predictive correction, and run-time control within the same block, and we use Orthogonal Novelty Transport (ONT) to govern slow-memory writes. We evaluate a 158M-parameter model in three stages spanning base language modeling, mathematical continuation, and 4096-token continuation. Removing mHC raises the Stage-A final LM loss from 12.630 to 15.127, while adaptive sparse control improves the Stage-B final LM loss from 12.137 to 10.787 relative to a matched fixed-ratio continuation. The full route remains stable at sequence length 4096, where Stage C ends with final LM loss 11.582 and improves the delayed-identifier diagnostic from 14.396 to 12.031 in key cross-entropy. Taken together, these results show that long-context autoregressive modeling can be organized around a broader division of labor than attention alone.

The rise of generative AI is posing increasing risks to online information integrity and civic discourse. Most concretely, such risks can materialise in the form of mis- and disinformation. As a mitigation, media-literacy and transparency tools have been developed to address factuality of information and the reliability and ideological leaning of information sources. However, a subtler but possibly no less harmful threat to civic discourse is to use of persuasion or manipulation by exploiting human cognitive biases and related cognitive limitations. To the best of our knowledge, no tools exist to directly detect and mitigate the presence of triggers of such cognitive biases in online information. We present VIGIL (VIrtual GuardIan angeL), the first browser extension for real-time cognitive bias trigger detection and mitigation, providing in-situ scroll-synced detection, LLM-powered reformulation with full reversibility, and privacy-tiered inference from fully offline to cloud. VIGIL is built to be extensible with third-party plugins, with several plugins that are rigorously validated against NLP benchmarks are already included. It is open-sourced at https://github.com/aida-ugent/vigil.

Standard attention scales quadratically with sequence length. Efficient attention methods reduce this O(n^2) cost, but when retrofitted into pretrained models, they often degrade perplexity, downstream accuracy, or both. We introduce Focus, a method that learns which token pairs matter. Focus adds a small set of learnable centroids–as few as 148K parameters per layer–that act as gates: only token pairs belonging to the same centroid group attend to each other over long ranges. Focus is composable: it can be added to any pretrained model by training only the centroids while keeping all original weights frozen. Experiments show that composing Focus onto pretrained models yields zero degradation on downstream benchmarks across model sizes from 124M to 70B parameters and five attention architectures. Surprisingly, sparse Focus attention outperforms full attention at 124M scale (30.3 vs. 31.4 perplexity) and matches full attention when trained from scratch at 7B scale (13.82 vs. 13.89). Focus is also fast: top-k group membership gives a 2x speedup with better quality than the original pretrained model. Using our FlashAttention decomposition, Focus achieves an 8.6x speedup at 1M tokens without custom kernels.

Large language models (LLMs) have demonstrated impressive capabilities across various tasks, but the billion-scale parameters pose deployment challenges. Although existing methods attempt to reduce the scale of LLMs, they require either special hardware support or expensive post-training to maintain model quality. To facilitate efficient and affordable model slimming, we propose a novel training-free compression method for LLMs, named "SoLA", which leverages Soft activation sparsity and Low-rAnk decomposition. SoLA can identify and retain a minority of components significantly contributing to inference, while compressing the majority through low-rank decomposition, based on our analysis of the activation pattern in the feed-forward network (FFN) of modern LLMs. To alleviate the decomposition loss, SoLA is equipped with an adaptive component-wise low-rank allocation strategy to assign appropriate truncation positions for different weight matrices. We conduct extensive experiments on LLaMA-2-7B/13B/70B and Mistral-7B models across a variety of benchmarks. SoLA exhibits remarkable improvement in both language modeling and downstream task accuracy without post-training. For example, with a 30% compression rate on the LLaMA-2-70B model, SoLA surpasses the state-of-the-art method by reducing perplexity from 6.95 to 4.44 and enhancing downstream task accuracy by 10%.

We show that PLDR-LLMs pretrained at self-organized criticality exhibit reasoning at inference time. The characteristics of PLDR-LLM deductive outputs at criticality is similar to second-order phase transitions. At criticality, the correlation length diverges, and the deductive outputs attain a metastable steady state. The steady state behaviour suggests that deductive outputs learn representations equivalent to scaling functions, universality classes and renormalization groups from the training dataset, leading to generalization and reasoning capabilities in the process. We can then define an order parameter from the global statistics of the model's deductive output parameters at inference. The reasoning capabilities of a PLDR-LLM is better when its order parameter is close to zero at criticality. This observation is supported by the benchmark scores of the models trained at near-criticality and sub-criticality. Our results provide a self-contained explanation on how reasoning manifests in large language models, and the ability to reason can be quantified solely from global model parameter values of the deductive outputs at steady state, without any need for evaluation of curated benchmark datasets through inductive output for reasoning and comprehension.