Papers

Practitioners have access to an abundance of language models and prompting strategies for solving many language modeling tasks; yet prior work shows that modeling performance is highly sensitive to both choices. Classical machine learning ensembling techniques offer a principled approach: aggregate predictions from multiple sources to achieve better performance than any single one. However, applying ensembling to language models during decoding is challenging: naively aggregating next-token probabilities yields samples from a locally normalized, biased approximation of the generally intractable ensemble distribution over strings. In this work, we introduce a unified framework for composing $K$ language models into $f$-ensemble distributions for a wide range of functions $f\colon\mathbb{R}_{\geq 0}^{K}\to\mathbb{R}_{\geq 0}$. To sample from these distributions, we propose a byte-level sequential Monte Carlo (SMC) algorithm that operates in a shared character space, enabling ensembles of models with mismatching vocabularies and consistent sampling in the limit. We evaluate a family of $f$-ensembles across prompt and model combinations for various structured text generation tasks, highlighting the benefits of alternative aggregation strategies over traditional probability averaging, and showing that better posterior approximations can yield better ensemble performance.

Introspection is a foundational cognitive ability, but its mechanism is not well understood. Recent work has shown that AI models can introspect. We study the mechanism of this introspection. We first extensively replicate Lindsey (2025)'s thought injection detection paradigm in large open-source models. We show that introspection in these models is content-agnostic: models can detect that an anomaly occurred even when they cannot reliably identify its content. The models confabulate injected concepts that are high-frequency and concrete (e.g., "apple"). They also require fewer tokens to detect an injection than to guess the correct concept (with wrong guesses coming earlier). We argue that a content-agnostic introspective mechanism is consistent with leading theories in philosophy and psychology.

Word Sense Disambiguation (WSD) remains a key challenge in Natural Language Processing (NLP), especially when dealing with rare or domain-specific senses that are often misinterpreted. While modern high-parameter Large Language Models (LLMs) such as GPT-4-Turbo have shown state-of-the-art WSD performance, their computational and energy demands limit scalability. This study investigates whether low-parameter LLMs (<4B parameters) can achieve comparable results through fine-tuning strategies that emphasize reasoning-driven sense identification. Using the FEWS dataset augmented with semi-automated, rationale-rich annotations, we fine-tune eight small-scale open-source LLMs (e.g. Gemma and Qwen). Our results reveal that Chain-of-Thought (CoT)-based reasoning combined with neighbour-word analysis achieves performance comparable to GPT-4-Turbo in zero-shot settings. Importantly, Gemma-3-4B and Qwen-3-4B models consistently outperform all medium-parameter baselines and state-of-the-art models on FEWS, with robust generalization to unseen senses. Furthermore, evaluation on the unseen "Fool Me If You Can” dataset confirms strong cross-domain adaptability without task-specific fine-tuning. This work demonstrates that with carefully crafted reasoning-centric fine-tuning, low-parameter LLMs can deliver accurate WSD while substantially reducing computational and energy demands.

Transformer architectures serve as the backbone for most modern Large Language Models, therefore their pretraining stability and convergence speed are of central concern. Motivated by the logical dependency of sequentially stacked layers, we propose Progressive Residual Warmup (ProRes) for language model pretraining. ProRes implements an "early layer learns first" philosophy by multiplying each layer's residual with a scalar that gradually warms up from 0 to 1, with deeper layers taking longer warmup steps. In this way, deeper layers wait for early layers to settle into a more stable regime before contributing to learning. We demonstrate the effectiveness of ProRes through pretraining experiments across various model scales, as well as normalization and initialization schemes. Comprehensive analysis shows that ProRes not only stabilizes pretraining but also introduces a unique optimization trajectory, leading to faster convergence, stronger generalization and better downstream performance. Our code is available at https://github.com/dandingsky/ProRes.

Test-time adaptation offers a promising avenue for improving reasoning performance in large language models without additional supervision, but existing approaches often apply a uniform optimization objective across all inputs, leading to inefficient or unstable adaptation on heterogeneous reasoning problems. We propose DiSCTT, a difficulty-aware, consensus-guided self-curriculum framework that dynamically allocates test-time optimization strategies based on instance-level epistemic uncertainty estimated from agreement among sampled reasoning trajectories. Inputs with high consensus are consolidated via supervised fine-tuning using majority-agreed solutions as pseudo-labels, while low-consensus inputs are optimized via reinforcement learning with a consensus-regularized objective that encourages diversity under relevance constraints. Across a broad suite of mathematical and general reasoning benchmarks, DiSCTT consistently outperforms strong test-time adaptation baselines, achieving higher accuracy with reduced variance and substantially lower computation and wall-clock training times. These results demonstrate that explicitly accounting for instance difficulty and uncertainty enables more stable, efficient, and effective test-time adaptation for reasoning models.

Model merging is a scalable alternative to multi-task training that combines the capabilities of multiple specialised models into a single model. This is particularly attractive for large speech foundation models, which are typically adapted through domain-specific fine-tuning, resulting in multiple customised checkpoints, for which repeating full fine-tuning when new data becomes available is computationally prohibitive. In this work, we study model merging for multi-domain ASR and benchmark 11 merging algorithms for 10 European Portuguese domains, evaluating in-domain accuracy, robustness under distribution shift, as well as English and multilingual performance. We further propose BoostedTSV-M, a new merging algorithm based on TSV-M that mitigates rank collapse via singular-value boosting and improves numerical stability. Overall, our approach outperforms full fine-tuning on European Portuguese while preserving out-of-distribution generalisation in a single model.

Being able to understand information is a key factor for a self-determined life and society. It is also very important for participating in democratic processes. The study of automatic text simplification is often limited by the availability of high quality material for the training and evaluation on automatic simplifiers. This is true for English, but more so for less resourced languages like Spanish, Catalan and Italian. In order to fill this gap, we present a corpus of original texts for these 3 languages, with high quality simplification produced by human experts in text simplification. It was developed within the iDEM project to assess the impact of Easy-to-Read (E2R) language for democratic participation. The original texts were compiled from domains related to this topic. The corpus includes different text types, selected based on relevance, copyright availability, and ethical standards. All texts were simplified to E2R level. The corpus is particularity valuable because it includes the first annotated corpus of its kind for the Catalan language. It also represents a noteworthy contribution for Spanish and Italian, offering high-quality, human-annotated language resources that are rarely available in these domains. The corpus will be made freely accessible to the public.

Punctuation restoration is essential for improving the readability and downstream utility of automatic speech recognition (ASR) outputs, yet remains underexplored for Persian despite its importance. We introduce PersianPunc, a large-scale, high-quality dataset of 17 million samples for Persian punctuation restoration, constructed through systematic aggregation and filtering of existing textual resources. We formulate punctuation restoration as a token-level sequence labeling task and fine-tune ParsBERT to achieve strong performance. Through comparative evaluation, we demonstrate that while large language models can perform punctuation restoration, they suffer from critical limitations: over-correction tendencies that introduce undesired edits beyond punctuation insertion (particularly problematic for speech-to-text pipelines) and substantially higher computational requirements. Our lightweight BERT-based approach achieves a macro-averaged F1 score of 91.33% on our test set while maintaining efficiency suitable for real-time applications. We make our dataset (https://huggingface.co/datasets/MohammadJRanjbar/persian-punctuation-restoration) and model (https://huggingface.co/MohammadJRanjbar/parsbert-persian-punctuation) publicly available to facilitate future research in Persian NLP and provide a scalable framework applicable to other morphologically rich, low-resource languages.

Assessing whether an article supports an assertion is essential for hallucination detection and claim verification. While large language models (LLMs) have the potential to automate this task, achieving strong performance requires frontier models such as GPT-5 that are prohibitively expensive to deploy at scale. To efficiently perform biomedical evidence attribution, we present Med-V1, a family of small language models with only three billion parameters. Trained on high-quality synthetic data newly developed in this study, Med-V1 substantially outperforms (+27.0% to +71.3%) its base models on five biomedical benchmarks unified into a verification format. Despite its smaller size, Med-V1 performs comparably to frontier LLMs such as GPT-5, along with high-quality explanations for its predictions. We use Med-V1 to conduct a first-of-its-kind use case study that quantifies hallucinations in LLM-generated answers under different citation instructions. Results show that the format instruction strongly affects citation validity and hallucination, with GPT-5 generating more claims but exhibiting hallucination rates similar to GPT-4o. Additionally, we present a second use case showing that Med-V1 can automatically identify high-stakes evidence misattributions in clinical practice guidelines, revealing potentially negative public health impacts that are otherwise challenging to identify at scale. Overall, Med-V1 provides an efficient and accurate lightweight alternative to frontier LLMs for practical, real-world biomedical evidence attribution. Med-V1 is available at https://github.com/NLM-DIR/Med-V1.

Large language models show that simple autoregressive training can yield scalable and coherent generation, but extending this paradigm to speech remains challenging due to the entanglement of semantic and acoustic information. Most existing speech language models rely on text supervision, hierarchical token streams, or complex hybrid architectures, departing from the single-stream generative pretraining paradigm that has proven effective in text. In this work, we introduce WavSLM, a speech language model trained by quantizing and distilling self-supervised WavLM representations into a single codebook and optimizing an autoregressive next-chunk prediction objective. WavSLM jointly models semantic and acoustic information within a single token stream without text supervision or text pretraining. Despite its simplicity, it achieves competitive performance on consistency benchmarks and speech generation while using fewer parameters, less training data, and supporting streaming inference.

AI safety via debate and reinforcement learning from AI feedback (RLAIF) are both proposed methods for scalable oversight of advanced AI systems, yet no formal framework relates them or characterizes when debate offers an advantage. We analyze this by parameterizing debate's value through the geometry of knowledge divergence between debating models. Using principal angles between models' representation subspaces, we prove that the debate advantage admits an exact closed form. When models share identical training corpora, debate reduces to RLAIF-like where a single-agent method recovers the same optimum. When models possess divergent knowledge, debate advantage scales with a phase transition from quadratic regime (debate offers negligible benefit) to linear regime (debate is essential). We classify three regimes of knowledge divergence (shared, one-sided, and compositional) and provide existence results showing that debate can achieve outcomes inaccessible to either model alone, alongside a negative result showing that sufficiently strong adversarial incentives cause coordination failure in the compositional regime, with a sharp threshold separating effective from ineffective debate. We offer the first formal connection between debate and RLAIF, a geometric foundation for understanding when adversarial oversight protocols are justified, and connection to the problem of eliciting latent knowledge across models with complementary information.

Multimodal sarcasm detection requires resolving pragmatic incongruity across textual, acoustic, and visual cues through cross-modal reasoning. To enable robust sarcasm reasoning with foundation models, we propose SarcasmMiner, a reinforcement learning based post-training framework that resists hallucination in multimodal reasoning. We reformulate sarcasm detection as structured reasoning and adopt a dual-track distillation strategy: high-quality teacher trajectories initialize the student model, while the full set of trajectories trains a generative reward model (GenRM) to evaluate reasoning quality. The student is optimized with group relative policy optimization (GRPO) using decoupled rewards for accuracy and reasoning quality. On MUStARD++, SarcasmMiner increases F1 from 59.83% (zero-shot), 68.23% (supervised finetuning) to 70.22%. These findings suggest that reasoning-aware reward modeling enhances both performance and multimodal grounding.

We present the Multilingual Cloud Corpus, the first national-scale, parallel, multimodal linguistic dataset of Bangladesh's ethnic and indigenous languages. Despite being home to approximately 40 minority languages spanning four language families, Bangladesh has lacked a systematic, cross-family digital corpus for these predominantly oral, computationally "zero resource" varieties, 14 of which are classified as endangered. Our corpus comprises 85792 structured textual entries, each containing a Bengali stimulus text, an English translation, and an IPA transcription, together with approximately 107 hours of transcribed audio recordings, covering 42 language varieties from the Tibeto-Burman, Indo-European, Austro-Asiatic, and Dravidian families, plus two genetically unclassified languages. The data were collected through systematic fieldwork over 90 days across nine districts of Bangladesh, involving 16 data collectors, 77 speakers, and 43 validators, following a predefined elicitation template of 2224 unique items organized at three levels of linguistic granularity: isolated lexical items (475 words across 22 semantic domains), grammatical constructions (887 sentences across 21 categories including verbal conjugation paradigms), and directed speech (862 prompts across 46 conversational scenarios). Post-field processing included IPA transcription by 10 linguists with independent adjudication by 6 reviewers. The complete dataset is publicly accessible through the Multilingual Cloud platform (multiling.cloud), providing searchable access to annotated audio and textual data for all documented varieties. We describe the corpus design, fieldwork methodology, dataset structure, and per-language coverage, and discuss implications for endangered language documentation, low-resource NLP, and digital preservation in linguistically diverse developing countries.

VietJobs is the first large-scale, publicly available corpus of Vietnamese job advertisements, comprising 48,092 postings and over 15 million words collected from all 34 provinces and municipalities across Vietnam. The dataset provides extensive linguistic and structured information, including job titles, categories, salaries, skills, and employment conditions, covering 16 occupational domains and multiple employment types (full-time, part-time, and internship). Designed to support research in natural language processing and labour market analytics, VietJobs captures substantial linguistic, regional, and socio-economic diversity. We benchmark several generative large language models (LLMs) on two core tasks: job category classification and salary estimation. Instruction-tuned models such as Qwen2.5-7B-Instruct and Llama-SEA-LION-v3-8B-IT demonstrate notable gains under few-shot and fine-tuned settings, while highlighting challenges in multilingual and Vietnamese-specific modelling for structured labour market prediction. VietJobs establishes a new benchmark for Vietnamese NLP and offers a valuable foundation for future research on recruitment language, socio-economic representation, and AI-driven labour market analysis. All code and resources are available at: https://github.com/VinNLP/VietJobs.

Speculative decoding accelerates inference for Large Language Models by using a lightweight draft model to propose candidate tokens that are verified in parallel by a larger target model. Prior work shows that the draft model often dominates speculative decoding latency, since it generates tokens sequentially and incurs high cost from its language modeling head as vocabulary size grows. This exposes a fundamental trade-off in draft model design: larger vocabularies improve token coverage and agreement with the target model, but incur higher draft latency, while smaller vocabularies reduce latency at the risk of missing tokens required for accurate draft generation. We address this trade-off through vocabulary trimming for draft models, motivated by the observation that domain-specific workloads use only a small fraction of the full vocabulary. We cast draft vocabulary selection as a constrained optimization problem that balances token coverage and draft latency. Coverage is computed over assistant responses in the training data, while latency is estimated using architecture-aware FLOPs that capture the cost of the language modeling head as a function of vocabulary size. We optimize a utility function with a Tree-structured Parzen Estimator to efficiently explore the coverage-latency Pareto frontier under a minimum coverage constraint. Experiments show improved speculative decoding throughput while reducing draft vocabularies by up to 97% with high coverage. On domain-specific tasks, we achieve up to 16% latency reduction and 20% throughput improvement, and up to 6.7% throughput gains on diverse out-of-distribution tasks.

Retrieval-Augmented Generation (RAG) enhances large language models by incorporating external knowledge. However, existing vector-based methods often fail on global sensemaking tasks that require reasoning across many documents. GraphRAG addresses this by organizing documents into a knowledge graph with hierarchical communities that can be recursively summarized. Current GraphRAG approaches rely on Leiden clustering for community detection, but we prove that on sparse knowledge graphs, where average degree is constant and most nodes have low degree, modularity optimization admits exponentially many near-optimal partitions, making Leiden-based communities inherently non-reproducible. To address this, we propose replacing Leiden with k-core decomposition, which yields a deterministic, density-aware hierarchy in linear time. We introduce a set of lightweight heuristics that leverage the k-core hierarchy to construct size-bounded, connectivity-preserving communities for retrieval and summarization, along with a token-budget-aware sampling strategy that reduces LLM costs. We evaluate our methods on real-world datasets including financial earnings transcripts, news articles, and podcasts, using three LLMs for answer generation and five independent LLM judges for head-to-head evaluation. Across datasets and models, our approach consistently improves answer comprehensiveness and diversity while reducing token usage, demonstrating that k-core-based GraphRAG is an effective and efficient framework for global sensemaking.

We introduce a framework for learning continuous neural representations of formal specifications by distilling the geometry of their semantics into a latent space. Existing approaches rely either on symbolic kernels – which preserve behavioural semantics but are computationally prohibitive, anchor-dependent, and non-invertible – or on syntax-based neural embeddings that fail to capture underlying structures. Our method bridges this gap: using a teacher-student setup, we distill a symbolic robustness kernel into a Transformer encoder. Unlike standard contrastive methods, we supervise the model with a continuous, kernel-weighted geometric alignment objective that penalizes errors in proportion to their semantic discrepancies. Once trained, the encoder produces embeddings in a single forward pass, effectively mimicking the kernel's logic at a fraction of its computational cost. We apply our framework to Signal Temporal Logic (STL), demonstrating that the resulting neural representations faithfully preserve the semantic similarity of STL formulae, accurately predict robustness and constraint satisfaction, and remain intrinsically invertible. Our proposed approach enables highly efficient, scalable neuro-symbolic reasoning and formula reconstruction without repeated kernel computation at runtime.

Diffusion LLMs have been proposed as an alternative to autoregressive LLMs. Curiously, they are especially capable if the generation length, i.e., the number of tokens the model has to output, is set to a much higher value than the correct answer length, and the model pads its answer with end-of-sequence (EoS) tokens. We hypothesize that off-the-shelf masked diffusion LLMs use the representations of EoS tokens as additional computing capacity, which enhances their performance. We experiment with the diffusion models LLaDA1.5, LLaDA2.0-mini, and Dream-v0 on three reasoning tasks: Addition, Entity Tracking, and Sudoku. In a controlled prompting experiment, we confirm that adding EoS tokens improves the LLMs' performance. To further verify whether their representations are used for hidden computations, we perform a causal intervention and transfer the hidden states of the EoS tokens between generations, which increases the models' relative likelihood of outputting the counterfactual answer. The behavioral experiments and the causal interventions indicate that fully bidirectional masked diffusion LLMs can indeed perform latent reasoning in the representations of EoS tokens. Furthermore, we find that these results generalize beyond toy tasks and that providing the model with additional EoS tokens also improves performance on GSM8K and two-hop reasoning.

Modern language models define distributions over strings, but downstream tasks often require different output formats. For instance, a model that generates byte-pair strings does not directly produce word-level predictions, and a DNA model does not directly produce amino-acid sequences. In such cases, a deterministic string-to-string transformation can convert the model's output to the desired form. This is a familiar pattern in probability theory: applying a function $f$ to a random variable $X\sim p$ yields a transformed random variable $f(X)$ with an induced distribution. While such transformations are occasionally used in language modeling, prior work does not treat them as yielding new, fully functional language models. We formalize this perspective and introduce a general framework for language models derived from deterministic string-to-string transformations. We focus on transformations representable as finite-state transducers – a commonly used state-machine abstraction for efficient string-to-string mappings. We develop algorithms that compose a language model with an FST to *marginalize* over source strings mapping to a given target, propagating probabilities through the transducer without altering model parameters and enabling *conditioning* on transformed outputs. We present an exact algorithm, an efficient approximation, and a theoretical analysis. We conduct experiments in three domains: converting language models from tokens to bytes, from tokens to words, and from DNA to amino acids. These experiments demonstrate inference-time adaptation of pretrained language models to match application-specific output requirements.

This Guideline presents a systematic and operationalizable annotation framework for representing legal argumentation structures in judicial decisions. Grounded in theories of legal reasoning and argumentation, the framework aims to reveal the logical organization of judicial reasoning and provide a reliable foundation for computational analysis. At the element level, the Guideline distinguishes between the non-propositional layer and the propositional layer. The non-propositional layer consists of two elements: Issue and Non-argumentative Component. At the propositional level, the Guideline defines four proposition types: General Normative Judgment, Particular Normative Judgment, General Factual Judgment, and Particular Factual Judgment. At the relational level, five relation types are defined to represent argumentative structures: Support, Attack, Joint, Match, and Identity. These relations capture positive and negative argumentative connections, conjunctive reasoning structures, correspondences between legal norms and case facts, and identity or semantic equivalence between propositions. The Guideline further specifies formal representation rules and visualization conventions for both basic and nested structures, enabling consistent visualization of complex argumentation patterns. In addition, it establishes a standardized annotation workflow and consistency control mechanisms to ensure the reproducibility and reliability of annotated data. By providing a clear conceptual model, formal representation rules, and practical annotation procedures, this Guideline supports large-scale analysis of judicial reasoning and future research in legal argument mining, computational modeling of legal reasoning, and AI-assisted legal analysis.

Semi-structured N:M sparsity and low-bit quantization (e.g., 1.58-bit BitNet) are two promising approaches for improving the efficiency of large language models (LLMs), yet they have largely been studied in isolation. In this work, we investigate their interaction and show that 1.58-bit BitNet is naturally more compatible with N:M sparsity than full-precision models. To study this effect, we propose Sparse-BitNet, a unified framework that jointly applies 1.58-bit quantization and dynamic N:M sparsification while ensuring stable training for the first time. Across multiple model scales and training regimes (sparse pretraining and dense-to-sparse schedules), 1.58-bit BitNet consistently exhibits smaller performance degradation than full-precision baselines at the same sparsity levels and can tolerate higher structured sparsity before accuracy collapse. Moreover, using our custom sparse tensor core, Sparse-BitNet achieves substantial speedups in both training and inference, reaching up to 1.30X. These results highlight that combining extremely low-bit quantization with semi-structured N:M sparsity is a promising direction for efficient LLMs. Code available at https://github.com/AAzdi/Sparse-BitNet

Large language models (LLMs) are increasingly used as judges of chain-of-thought (CoT) reasoning, yet it remains unclear whether they can reliably assess process faithfulness rather than merely answer plausibility. We introduce C2-Faith, a benchmark built from PRM800K that explicitly decomposes faithfulness into two complementary dimensions: causality (whether each step logically follows from prior context) and coverage (whether essential intermediate inferences are present). Using controlled perturbations, we construct examples with known causal error positions by replacing a single step with a logically inconsistent variant, and with controlled coverage deletions at varying rates, enabling direct measurement against reference labels. We evaluate three frontier LLM judges across three tasks: binary causal detection, causal step localization, and coverage scoring. Our results reveal that judge reliability is highly task-dependent, with no single model dominating across settings. While models often detect that an error exists, they struggle to accurately localize it, indicating a substantial gap between detection and attribution. Moreover, all judges systematically overestimate reasoning completeness, assigning high coverage scores even when substantial portions of intermediate reasoning are missing. These findings expose fundamental limitations of LLM judges in process-level evaluation and highlight the need for more reliable and calibrated methods when using LLMs to assess reasoning quality.

Understanding reasoning in large language models is complicated by evaluations that conflate multiple reasoning types. We isolate analogical reasoning, where a model transfers an attribute between entities that share known properties, and study when such transfer can emerge from training. To make the problem analytically tractable, we study a minimal transformer-style abstraction that isolates how learned representations support analogical reasoning. Within this setting, we prove three key results. First, joint training on similarity and attribution premises enables analogical reasoning through aligned representations. Second, sequential training succeeds only when similarity structure is learned before specific attributes, revealing a curriculum asymmetry. Third, in our stylized setting, two-hop reasoning $(a \to b, b \to c \Rightarrow a \to c)$ can be viewed as analogical reasoning with identity bridges $(b=b)$, which appear explicitly in training data. Together, these results reveal a unified mechanism: entities with shared properties become aligned in representation space, enabling property transfer through feature resemblance. Experiments with architectures up to 8B parameters show qualitative agreement with the theory and suggest that representational geometry plays an important role in analogical reasoning beyond the stylized model.

This paper introduces a new dimension for validating algorithmic decisions about humans by measuring the fidelity of their representations. Representation Fidelity measures if decisions about a person rest on reasonable grounds. We propose to operationalize this notion by measuring the distance between two representations of the same person: (1) an externally prescribed input representation on which the decision is based, and (2) a self-description provided by the human subject of the decision, used solely to validate the input representation. We examine the nature of discrepancies between these representations, how such discrepancies can be quantified, and derive a generic typology of representation mismatches that determine the degree of representation fidelity. We further present the first benchmark for evaluating representation fidelity based on a dataset of loan-granting decisions. Our Loan-Granting Self-Representations Corpus 2025 consists of a large corpus of 30 000 synthetic natural language self-descriptions derived from corresponding representations of applicants in the German Credit Dataset, along with expert annotations of representation mismatches between each pair of representations.

The growing scale of ad auctions on online advertising platforms has intensified competition, making manual bidding impractical and necessitating auto-bidding to help advertisers achieve their economic goals. Current auto-bidding methods have evolved to use offline reinforcement learning or generative methods to optimize bidding strategies, but they can sometimes behave counterintuitively due to the black-box training manner and limited mode coverage of datasets, leading to challenges in understanding task status and generalization in dynamic ad environments. Large language models (LLMs) offer a promising solution by leveraging prior human knowledge and reasoning abilities to improve auto-bidding performance. However, directly applying LLMs to auto-bidding faces difficulties due to the need for precise actions in competitive auctions and the lack of specialized auto-bidding knowledge, which can lead to hallucinations and suboptimal decisions. To address these challenges, we propose a hierarchical Large autoBidding Model (LBM) to leverage the reasoning capabilities of LLMs for developing a superior auto-bidding strategy. This includes a high-level LBM-Think model for reasoning and a low-level LBM-Act model for action generation. Specifically, we propose a dual embedding mechanism to efficiently fuse two modalities, including language and numerical inputs, for language-guided training of the LBM-Act; then, we propose an offline reinforcement fine-tuning technique termed GQPO for mitigating the LLM-Think's hallucinations and enhancing decision-making performance without simulation or real-world rollout like previous multi-turn LLM-based methods. Experiments demonstrate the superiority of a generative backbone based on our LBM, especially in an efficient training manner and generalization ability.

Speech Large Language Models route speech encoder representations into an LLM decoder that typically accounts for over 90% of total parameters. We study how much of this decoder capacity is actually needed for speech tasks. Across two LLM families and three scales (1-8B), we show that decoder redundancy is largely inherited from the pretrained LLM: text and speech inputs yield similar redundant blocks. We then measure excess capacity by pruning decoder layers and analysing post-pruning healing to increase robustness. Our findings show that 7-8B models retain good ASR performance with only 60% of decoder layers, and the same trend extends to smaller scales with reduced pruning tolerance. We then generalise to speech translation, and show that the same blocks of layers are redundant across speech encoders, tasks and languages, indicating that a more global redundancy structure exists, enabling a single pruned and multi-tasks SpeechLLM backbone to be deployed.

The Abstraction and Reasoning Corpus (ARC-AGI) probes few-shot abstraction and rule induction on small visual grids, but progress is difficult to measure on static collections of hand-authored puzzles due to overfitting, dataset leakage, and memorisation. We introduce ARC-TGI (ARC Task Generators Inventory), an open-source framework for task-family generators: compact Python programs that sample diverse ARC-AGI tasks while preserving a latent rule. ARC-TGI is built around a solver-facing representation: each generated task is paired with natural-language input and transformation reasoning chains and partially evaluated Python code implementing sampling, transformation, and episode construction. Crucially, ARC-TGI supports task-level constraints so that training examples collectively expose the variations needed to infer the underlying rule, a requirement for human-solvable ARC tasks that independent per-example sampling often fails to guarantee. All generators undergo human refinement and local verification to keep both grids and reasoning traces natural and consistent under variation. We release 461 generators covering 180 ARC-Mini tasks, 215 ARC-AGI-1 tasks (200 train, 15 test), and 66 ARC-AGI-2 tasks (55 train, 11 test), enabling scalable dataset sampling and controlled benchmarking.

Time series forecasting has witnessed an increasing demand across diverse industrial applications, where accurate predictions are pivotal for informed decision-making. Beyond numerical time series data, reliable forecasting in practical scenarios requires integrating diverse exogenous factors. Such exogenous information is often multi-dimensional or even multimodal, introducing heterogeneous interactions that unimodal time series models struggle to capture. In this paper, we delve into an aviation maintenance scenario and identify three distinct types of exogenous factors that influence temporal dynamics through distinct interaction modes. Based on this empirical insight, we propose Aura, a universal framework that explicitly organizes and encodes heterogeneous external information according to its interaction mode with the target time series. Specifically, Aura utilizes a tailored tripartite encoding mechanism to embed heterogeneous features into well-established time series models, ensuring seamless integration of non-sequential context. Extensive experiments on a large-scale, three-year industrial dataset from China Southern Airlines, covering the Boeing 777 and Airbus A320 fleets, demonstrate that Aura consistently achieves state-of-the-art performance across all baselines and exhibits superior adaptability. Our findings highlight Aura's potential as a general-purpose enhancement for aviation safety and reliability.

Urdu toxic span detection remains limited because most existing systems rely on sentence-level classification and fail to identify the specific toxic spans within those text. It is further exacerbated by the multiple factors i.e. lack of token-level annotated resources, linguistic complexity of Urdu, frequent code-switching, informal expressions, and rich morphological variations. In this research, we propose MUTEX: a multilingual transformer combined with conditional random fields (CRF) for Urdu toxic span detection framework that uses manually annotated token-level toxic span dataset to improve performance and interpretability. MUTEX uses XLM RoBERTa with CRF layer to perform sequence labeling and is tested on multi-domain data extracted from social media, online news, and YouTube reviews using token-level F1 to evaluate fine-grained span detection. The results indicate that MUTEX achieves 60% token-level F1 score that is the first supervised baseline for Urdu toxic span detection. Further examination reveals that transformer-based models are more effective at implicitly capturing the contextual toxicity and are able to address the issues of code-switching and morphological variation than other models.

Extending large language models to low-resource languages is essential for global accessibility, but training separate models per language is prohibitively expensive. Mixture-of-Experts (MoE) architectures address this by adding sparse language-specific parameters, but determining how many experts each layer needs remains an open question. Current approaches allocate experts based on layer-level similarity, yet language processing exhibits fine-grained specialization at individual neurons. We propose $\textbf{NeuronMoE}$, a method that analyzes language-specific neurons across all transformer components to guide expert allocation per layer based on empirically measured cross-lingual neuron diversity. Applied to Llama-3.2-3B for low-resource languages (Greek, Turkish, and Hungarian), this approach achieves approximately 40% average parameter reduction while matching the performance of the LayerMoE baseline. We find that low-resource language experts independently develop neuron specialization patterns mirroring the high-resource language, which are concentrated in early and late layers. This reveals potential universal architectural principles in how multilingual models organize linguistic knowledge.