2026-02-06 | | Total: 101
Autoregressive large language models (LLMs) deliver strong performance but require inherently sequential decoding, leading to high inference latency and poor GPU utilization. Speculative decoding mitigates this bottleneck by using a fast draft model whose outputs are verified in parallel by the target LLM; however, existing methods still rely on autoregressive drafting, which remains sequential and limits practical speedups. Diffusion LLMs offer a promising alternative by enabling parallel generation, but current diffusion models typically underperform compared with autoregressive models. In this paper, we introduce DFlash, a speculative decoding framework that employs a lightweight block diffusion model for parallel drafting. By generating draft tokens in a single forward pass and conditioning the draft model on context features extracted from the target model, DFlash enables efficient drafting with high-quality outputs and higher acceptance rates. Experiments show that DFlash achieves over 6x lossless acceleration across a range of models and tasks, delivering up to 2.5x higher speedup than the state-of-the-art speculative decoding method EAGLE-3.
Memory is increasingly central to Large Language Model (LLM) agents operating beyond a single context window, yet most existing systems rely on offline, query-agnostic memory construction that can be inefficient and may discard query-critical information. Although runtime memory utilization is a natural alternative, prior work often incurs substantial overhead and offers limited explicit control over the performance-cost trade-off. In this work, we present \textbf{BudgetMem}, a runtime agent memory framework for explicit, query-aware performance-cost control. BudgetMem structures memory processing as a set of memory modules, each offered in three budget tiers (i.e., \textsc{Low}/\textsc{Mid}/\textsc{High}). A lightweight router performs budget-tier routing across modules to balance task performance and memory construction cost, which is implemented as a compact neural policy trained with reinforcement learning. Using BudgetMem as a unified testbed, we study three complementary strategies for realizing budget tiers: implementation (method complexity), reasoning (inference behavior), and capacity (module model size). Across LoCoMo, LongMemEval, and HotpotQA, BudgetMem surpasses strong baselines when performance is prioritized (i.e., high-budget setting), and delivers better accuracy-cost frontiers under tighter budgets. Moreover, our analysis disentangles the strengths and weaknesses of different tiering strategies, clarifying when each axis delivers the most favorable trade-offs under varying budget regimes.
Existing techniques for accelerating language model inference, such as speculative decoding, require training auxiliary speculator models and building and deploying complex inference pipelines. We consider a new approach for converting a pretrained autoregressive language model from a slow single next token prediction model into a fast standalone multi-token prediction model using a simple online distillation objective. The final model retains the exact same implementation as the pretrained initial checkpoint and is deployable without the addition of any auxiliary verifier or other specialized inference code. On GSM8K, our method produces models that can decode more than $3\times$ faster on average at $<5\%$ drop in accuracy relative to single token decoding performance.
Large language models (LLMs) are increasingly being used in a zero-shot fashion to assess mental health conditions, yet we have limited knowledge on what factors affect their accuracy. In this study, we utilize a clinical dataset of natural language narratives and self-reported PTSD severity scores from 1,437 individuals to comprehensively evaluate the performance of 11 state-of-the-art LLMs. To understand the factors affecting accuracy, we systematically varied (i) contextual knowledge like subscale definitions, distribution summary, and interview questions, and (ii) modeling strategies including zero-shot vs few shot, amount of reasoning effort, model sizes, structured subscales vs direct scalar prediction, output rescaling and nine ensemble methods. Our findings indicate that (a) LLMs are most accurate when provided with detailed construct definitions and context of the narrative; (b) increased reasoning effort leads to better estimation accuracy; (c) performance of open-weight models (Llama, Deepseek), plateau beyond 70B parameters while closed-weight (o3-mini, gpt-5) models improve with newer generations; and (d) best performance is achieved when ensembling a supervised model with the zero-shot LLMs. Taken together, the results suggest choice of contextual knowledge and modeling strategies is important for deploying LLMs to accurately assess mental health.
Diffusion large language models (dLLMs) have emerged as a promising alternative for text generation, distinguished by their native support for parallel decoding. In practice, block inference is crucial for avoiding order misalignment in global bidirectional decoding and improving output quality. However, the widely-used fixed, predefined block (naive) schedule is agnostic to semantic difficulty, making it a suboptimal strategy for both quality and efficiency: it can force premature commitments to uncertain positions while delaying easy positions near block boundaries. In this work, we analyze the limitations of naive block scheduling and disclose the importance of dynamically adapting the schedule to semantic difficulty for reliable and efficient inference. Motivated by this, we propose Dynamic Sliding Block (DSB), a training-free block scheduling method that uses a sliding block with a dynamic size to overcome the rigidity of the naive block. To further improve efficiency, we introduce DSB Cache, a training-free KV-cache mechanism tailored to DSB. Extensive experiments across multiple models and benchmarks demonstrate that DSB, together with DSB Cache, consistently improves both generation quality and inference efficiency for dLLMs. Code is released at https://github.com/lizhuo-luo/DSB.
Semantic representations can be framed as a structured, dynamic knowledge space through which humans navigate to retrieve and manipulate meaning. To investigate how humans traverse this geometry, we introduce a framework that represents concept production as navigation through embedding space. Using different transformer text embedding models, we construct participant-specific semantic trajectories based on cumulative embeddings and extract geometric and dynamical metrics, including distance to next, distance to centroid, entropy, velocity, and acceleration. These measures capture both scalar and directional aspects of semantic navigation, providing a computationally grounded view of semantic representation search as movement in a geometric space. We evaluate the framework on four datasets across different languages, spanning different property generation tasks: Neurodegenerative, Swear verbal fluency, Property listing task in Italian, and in German. Across these contexts, our approach distinguishes between clinical groups and concept types, offering a mathematical framework that requires minimal human intervention compared to typical labor-intensive linguistic pre-processing methods. Comparison with a non-cumulative approach reveals that cumulative embeddings work best for longer trajectories, whereas shorter ones may provide too little context, favoring the non-cumulative alternative. Critically, different embedding models yielded similar results, highlighting similarities between different learned representations despite different training pipelines. By framing semantic navigation as a structured trajectory through embedding space, bridging cognitive modeling with learned representation, thereby establishing a pipeline for quantifying semantic representation dynamics with applications in clinical research, cross-linguistic analysis, and the assessment of artificial cognition.
Long reasoning models often struggle in multilingual settings: they tend to reason in English for non-English questions; when constrained to reasoning in the question language, accuracies drop substantially. The struggle is caused by the limited abilities for both multilingual question understanding and multilingual reasoning. To address both problems, we propose TRIT (Translation-Reasoning Integrated Training), a self-improving framework that integrates the training of translation into multilingual reasoning. Without external feedback or additional multilingual data, our method jointly enhances multilingual question understanding and response generation. On MMATH, our method outperforms multiple baselines by an average of 7 percentage points, improving both answer correctness and language consistency. Further analysis reveals that integrating translation training improves cross-lingual question alignment by over 10 percentage points and enhances translation quality for both mathematical questions and general-domain text, with gains up to 8.4 COMET points on FLORES-200.
Multiple-Choice Questions (MCQs) are often used to assess knowledge, reasoning abilities, and even values encoded in large language models (LLMs). While the effect of multilingualism has been studied on LLM factual recall, this paper seeks to investigate the less explored question of language-induced variation in value-laden MCQ responses. Are multilingual LLMs consistent in their responses across languages, i.e. behave like theoretical polyglots, or do they answer value-laden MCQs depending on the language of the question, like a multitude of monolingual models expressing different values through a single model? We release a new corpus, the Multilingual European Value Survey (MEVS), which, unlike prior work relying on machine translation or ad hoc prompts, solely comprises human-translated survey questions aligned in 8 European languages. We administer a subset of those questions to over thirty multilingual LLMs of various sizes, manufacturers and alignment-fine-tuning status under comprehensive, controlled prompt variations including answer order, symbol type, and tail character. Our results show that while larger, instruction-tuned models display higher overall consistency, the robustness of their responses varies greatly across questions, with certain MCQs eliciting total agreement within and across models while others leave LLM answers split. Language-specific behavior seems to arise in all consistent, instruction-fine-tuned models, but only on certain questions, warranting a further study of the selective effect of preference fine-tuning.
Large language models rely on kv-caches to avoid redundant computation during autoregressive decoding, but as context length grows, reading and writing the cache can quickly saturate GPU memory bandwidth. Recent work has explored KV-cache compression, yet most approaches neglect the data-dependent nature of kv-caches and their variation across layers. We introduce KV-CoRE KV-cache Compressibility by Rank Evaluation), an SVD-based method for quantifying the data-dependent low-rank compressibility of kv-caches. KV-CoRE computes the optimal low-rank approximation under the Frobenius norm and, being gradient-free and incremental, enables efficient dataset-level, layer-wise evaluation. Using this method, we analyze multiple models and datasets spanning five English domains and sixteen languages, uncovering systematic patterns that link compressibility to model architecture, training data, and language coverage. As part of this analysis, we employ the Normalized Effective Rank as a metric of compressibility and show that it correlates strongly with performance degradation under compression. Our study establishes a principled evaluation framework and the first large-scale benchmark of kv-cache compressibility in LLMs, offering insights for dynamic, data-aware compression and data-centric model development.
Modeling latent character states is crucial for consistent and engaging role-playing (RP) with large language models (LLMs). Yet, existing prompting-based approaches mainly capture surface actions, often failing to track the latent states that drive interaction. We revisit finite-state machines (FSMs), long used in game design to model state transitions. While effective in small, well-specified state spaces, traditional hand-crafted, rule-based FSMs struggle to adapt to the open-ended semantic space of RP. To address this, we introduce Codified Finite-State Machines (CFSMs), a framework that automatically codifies textual character profiles into FSMs using LLM-based coding. CFSMs extract key states and transitions directly from the profile, producing interpretable structures that enforce character consistency. To further capture uncertainty and variability, we extend CFSMs into Codified Probabilistic Finite-State Machines (CPFSMs), where transitions are modeled as probability distributions over states. Through both synthetic evaluations and real-world RP scenarios in established artifacts, we demonstrate that CFSM and CPFSM outperform generally applied baselines, verifying effectiveness not only in structured tasks but also in open-ended stochastic state exploration.
As large language models become smaller and more efficient, small reasoning models (SRMs) are crucial for enabling chain-of-thought (CoT) reasoning in resource-constrained settings. However, they are prone to faithfulness hallucinations, especially in intermediate reasoning steps. Existing mitigation methods based on online reinforcement learning rely on outcome-based rewards or coarse-grained CoT evaluation, which can inadvertently reinforce unfaithful reasoning when the final answer is correct. To address these limitations, we propose Faithfulness-Aware Step-Level Reinforcement Learning (FaithRL), introducing step-level supervision via explicit faithfulness rewards from a process reward model, together with an implicit truncated resampling strategy that generates contrastive signals from faithful prefixes. Experiments across multiple SRMs and Open-Book QA benchmarks demonstrate that FaithRL consistently reduces hallucinations in both the CoT and final answers, leading to more faithful and reliable reasoning. Code is available at https://github.com/Easy195/FaithRL.
This report presents EuroLLM-22B, a large language model trained from scratch to support the needs of European citizens by covering all 24 official European Union languages and 11 additional languages. EuroLLM addresses the issue of European languages being underrepresented and underserved in existing open large language models. We provide a comprehensive overview of EuroLLM-22B's development, including tokenizer design, architectural specifications, data filtering, and training procedures. Across a broad set of multilingual benchmarks, EuroLLM-22B demonstrates strong performance in reasoning, instruction following, and translation, achieving results competitive with models of comparable size. To support future research, we release our base and instruction-tuned models, our multilingual web pretraining data and updated EuroBlocks instruction datasets, as well as our pre-training and evaluation codebases.
Hate speech detection is commonly framed as a direct binary classification problem despite being a composite concept defined through multiple interacting factors that vary across legal frameworks, platform policies, and annotation guidelines. As a result, supervised models often overfit dataset-specific definitions and exhibit limited robustness under domain shift and annotation noise. We introduce xList-Hate, a diagnostic framework that decomposes hate speech detection into a checklist of explicit, concept-level questions grounded in widely shared normative criteria. Each question is independently answered by a large language model (LLM), producing a binary diagnostic representation that captures hateful content features without directly predicting the final label. These diagnostic signals are then aggregated by a lightweight, fully interpretable decision tree, yielding transparent and auditable predictions. We evaluate it across multiple hate speech benchmarks and model families, comparing it against zero-shot LLM classification and in-domain supervised fine-tuning. While supervised methods typically maximize in-domain performance, we consistently improves cross-dataset robustness and relative performance under domain shift. In addition, qualitative analysis of disagreement cases provides evidence that the framework can be less sensitive to certain forms of annotation inconsistency and contextual ambiguity. Crucially, the approach enables fine-grained interpretability through explicit decision paths and factor-level analysis. Our results suggest that reframing hate speech detection as a diagnostic reasoning task, rather than a monolithic classification problem, provides a robust, explainable, and extensible alternative for content moderation.
The quadratic complexity of attention mechanisms poses a critical bottleneck for large language models processing long contexts. While dynamic sparse attention methods offer input-adaptive efficiency, they face fundamental trade-offs: requiring preprocessing, lacking global evaluation, violating query independence, or incurring high computational overhead. We present RRAttention, a novel dynamic sparse attention method that simultaneously achieves all desirable properties through a head \underline{r}ound-\underline{r}obin (RR) sampling strategy. By rotating query sampling positions across attention heads within each stride, RRAttention maintains query independence while enabling efficient global pattern discovery with stride-level aggregation. Our method reduces complexity from $O(L^2)$ to $O(L^2/S^2)$ and employs adaptive Top-$τ$ selection for optimal sparsity. Extensive experiments on natural language understanding (HELMET) and multimodal video comprehension (Video-MME) demonstrate that RRAttention recovers over 99\% of full attention performance while computing only half of the attention blocks, achieving 2.4$\times$ speedup at 128K context length and outperforming existing dynamic sparse attention methods.
The rapid advancement of Large Language Models (LLMs) has catalyzed the development of autonomous agents capable of navigating complex environments. However, existing evaluations primarily adopt a deductive paradigm, where agents execute tasks based on explicitly provided rules and static goals, often within limited planning horizons. Crucially, this neglects the inductive necessity for agents to discover latent transition laws from experience autonomously, which is the cornerstone for enabling agentic foresight and sustaining strategic coherence. To bridge this gap, we introduce OdysseyArena, which re-centers agent evaluation on long-horizon, active, and inductive interactions. We formalize and instantiate four primitives, translating abstract transition dynamics into concrete interactive environments. Building upon this, we establish OdysseyArena-Lite for standardized benchmarking, providing a set of 120 tasks to measure an agent's inductive efficiency and long-horizon discovery. Pushing further, we introduce OdysseyArena-Challenge to stress-test agent stability across extreme interaction horizons (e.g., > 200 steps). Extensive experiments on 15+ leading LLMs reveal that even frontier models exhibit a deficiency in inductive scenarios, identifying a critical bottleneck in the pursuit of autonomous discovery in complex environments. Our code and data are available at https://github.com/xufangzhi/Odyssey-Arena
Large language models (LLMs) have achieved strong performance in language-centric tasks. However, in agentic settings, LLMs often struggle to anticipate action consequences and adapt to environment dynamics, highlighting the need for world-modeling capabilities in LLM-based agents. We propose Reinforcement World Model Learning (RWML), a self-supervised method that learns action-conditioned world models for LLM-based agents on textual states using sim-to-real gap rewards. Our method aligns simulated next states produced by the model with realized next states observed from the environment, encouraging consistency between internal world simulations and actual environment dynamics in a pre-trained embedding space. Unlike next-state token prediction, which prioritizes token-level fidelity (i.e., reproducing exact wording) over semantic equivalence and can lead to model collapse, our method provides a more robust training signal and is empirically less susceptible to reward hacking than LLM-as-a-judge. We evaluate our method on ALFWorld and $τ^2$ Bench and observe significant gains over the base model, despite being entirely self-supervised. When combined with task-success rewards, our method outperforms direct task-success reward RL by 6.9 and 5.7 points on ALFWorld and $τ^2$ Bench respectively, while matching the performance of expert-data training.
LLM-based assistants have been widely popularised after the release of ChatGPT. Concerns have been raised about their misuse in academia, given the difficulty of distinguishing between human-written and generated text. To combat this, automated techniques have been developed and shown to be effective, to some extent. However, prior work suggests that these methods often falsely flag essays from non-native speakers as generated, due to their low perplexity extracted from an LLM, which is supposedly a key feature of the detectors. We revisit these statements two years later, specifically in the Czech language setting. We show that the perplexity of texts from non-native speakers of Czech is not lower than that of native speakers. We further examine detectors from three separate families and find no systematic bias against non-native speakers. Finally, we demonstrate that contemporary detectors operate effectively without relying on perplexity.
Reinforcement Learning has emerged as a key driver for LLM reasoning. This capability is equally pivotal in long-context scenarios--such as long-dialogue understanding and structured data analysis, where the challenge extends beyond consuming tokens to performing rigorous deduction. While existing efforts focus on data synthesis or architectural changes, recent work points out that relying solely on sparse, outcome-only rewards yields limited gains, as such coarse signals are often insufficient to effectively guide the complex long-context reasoning. To address this, we propose LongR, a unified framework that enhances long-context performance by integrating a dynamic "Think-and-Read" mechanism, which interleaves reasoning with document consultation, with a contextual density reward based on relative information gain to quantify the utility of the relevant documents. Empirically, LongR achieves a 9% gain on LongBench v2 and consistent improvements on RULER and InfiniteBench, demonstrating robust efficiency in navigating extensive contexts. Furthermore, LongR consistently enhances performance across diverse RL algorithms (e.g., DAPO, GSPO). Finally, we conduct in-depth analyses to investigate the impact of reasoning chain length on efficiency and the model's robustness against distractors.
Retrieval-augmented generation (RAG) has become a key paradigm for knowledge-intensive question answering. However, existing multi-hop RAG systems remain inefficient, as they alternate between retrieval and reasoning at each step, resulting in repeated LLM calls, high token consumption, and unstable entity grounding across hops. We propose CompactRAG, a simple yet effective framework that decouples offline corpus restructuring from online reasoning. In the offline stage, an LLM reads the corpus once and converts it into an atomic QA knowledge base, which represents knowledge as minimal, fine-grained question-answer pairs. In the online stage, complex queries are decomposed and carefully rewritten to preserve entity consistency, and are resolved through dense retrieval followed by RoBERTa-based answer extraction. Notably, during inference, the LLM is invoked only twice in total - once for sub-question decomposition and once for final answer synthesis - regardless of the number of reasoning hops. Experiments on HotpotQA, 2WikiMultiHopQA, and MuSiQue demonstrate that CompactRAG achieves competitive accuracy while substantially reducing token consumption compared to iterative RAG baselines, highlighting a cost-efficient and practical approach to multi-hop reasoning over large knowledge corpora. The implementation is available at GitHub.
Mixture-of-Experts (MoE) architectures are evolving towards finer granularity to improve parameter efficiency. However, existing MoE designs face an inherent trade-off between the granularity of expert specialization and hardware execution efficiency. We propose OmniMoE, a system-algorithm co-designed framework that pushes expert granularity to its logical extreme. OmniMoE introduces vector-level Atomic Experts, enabling scalable routing and execution within a single MoE layer, while retaining a shared dense MLP branch for general-purpose processing. Although this atomic design maximizes capacity, it poses severe challenges for routing complexity and memory access. To address these, OmniMoE adopts a system-algorithm co-design: (i) a Cartesian Product Router that decomposes the massive index space to reduce routing complexity from O(N) to O(sqrt(N)); and (ii) Expert-Centric Scheduling that inverts the execution order to turn scattered, memory-bound lookups into efficient dense matrix operations. Validated on seven benchmarks, OmniMoE (with 1.7B active parameters) achieves 50.9% zero-shot accuracy across seven benchmarks, outperforming coarse-grained (e.g., DeepSeekMoE) and fine-grained (e.g., PEER) baselines. Crucially, OmniMoE reduces inference latency from 73ms to 6.7ms (a 10.9-fold speedup) compared to PEER, demonstrating that massive-scale fine-grained MoE can be fast and accurate. Our code is open-sourced at https://github.com/flash-algo/omni-moe.
Multi-domain machine translation (MDMT) aims to build a unified model capable of translating content across diverse domains. Despite the impressive machine translation capabilities demonstrated by large language models (LLMs), domain adaptation still remains a challenge for LLMs. Existing MDMT methods such as in-context learning and parameter-efficient fine-tuning often suffer from domain shift, parameter interference and limited generalization. In this work, we propose a neuron-efficient fine-tuning framework for MDMT that identifies and updates consensus-aligned neurons within LLMs. These neurons are selected by maximizing the mutual information between neuron behavior and domain features, enabling LLMs to capture both generalizable translation patterns and domain-specific nuances. Our method then fine-tunes LLMs guided by these neurons, effectively mitigating parameter interference and domain-specific overfitting. Comprehensive experiments on three LLMs across ten German-English and Chinese-English translation domains evidence that our method consistently outperforms strong PEFT baselines on both seen and unseen domains, achieving state-of-the-art performance.
Inaccuracies in existing or generated clinical text may lead to serious adverse consequences, especially if it is a misdiagnosis or incorrect treatment suggestion. With Large Language Models (LLMs) increasingly being used across diverse healthcare applications, comprehensive evaluation through dedicated benchmarks is crucial. However, such datasets remain scarce, especially across diverse languages and contexts. In this paper, we introduce MedErrBench, the first multilingual benchmark for error detection, localization, and correction, developed under the guidance of experienced clinicians. Based on an expanded taxonomy of ten common error types, MedErrBench covers English, Arabic and Chinese, with natural clinical cases annotated and reviewed by domain experts. We assessed the performance of a range of general-purpose, language-specific, and medical-domain language models across all three tasks. Our results reveal notable performance gaps, particularly in non-English settings, highlighting the need for clinically grounded, language-aware systems. By making MedErrBench and our evaluation protocols publicly-available, we aim to advance multilingual clinical NLP to promote safer and more equitable AI-based healthcare globally. The dataset is available in the supplementary material. An anonymized version of the dataset is available at: https://github.com/congboma/MedErrBench.
We investigate how transformer models represent complex verb paradigms in Turkish and Modern Hebrew, concentrating on how tokenization strategies shape this ability. Using the Blackbird Language Matrices task on natural data, we show that for Turkish -- with its transparent morphological markers -- both monolingual and multilingual models succeed, either when tokenization is atomic or when it breaks words into small subword units. For Hebrew, instead, monolingual and multilingual models diverge. A multilingual model using character-level tokenization fails to capture the language non-concatenative morphology, but a monolingual model with morpheme-aware segmentation performs well. Performance improves on more synthetic datasets, in all models.
Large language models (LLMs) have advanced the development of personalized learning in education. However, their inherent generation mechanisms often produce homogeneous responses to identical prompts. This one-size-fits-all mechanism overlooks the substantial heterogeneity in students cognitive and psychological, thereby posing potential safety risks to vulnerable groups. Existing safety evaluations primarily rely on context-independent metrics such as factual accuracy, bias, or toxicity, which fail to capture the divergent harms that the same response might cause across different student attributes. To address this gap, we propose the concept of Student-Tailored Personalized Safety and construct CASTLE based on educational theories. This benchmark covers 15 educational safety risks and 14 student attributes, comprising 92,908 bilingual scenarios. We further design three evaluation metrics: Risk Sensitivity, measuring the model ability to detect risks; Emotional Empathy, evaluating the model capacity to recognize student states; and Student Alignment, assessing the match between model responses and student attributes. Experiments on 18 SOTA LLMs demonstrate that CASTLE poses a significant challenge: all models scored below an average safety rating of 2.3 out of 5, indicating substantial deficiencies in personalized safety assurance.
RL-based post-training with GRPO is widely used to improve large language models on individual reasoning tasks. However, real-world deployment requires reliable performance across diverse tasks. A straightforward multi-task adaptation of GRPO often leads to imbalanced outcomes, with some tasks dominating optimization while others stagnate. Moreover, tasks can vary widely in how frequently prompts yield zero advantages (and thus zero gradients), which further distorts their effective contribution to the optimization signal. To address these issues, we propose a novel Multi-Task GRPO (MT-GRPO) algorithm that (i) dynamically adapts task weights to explicitly optimize worst-task performance and promote balanced progress across tasks, and (ii) introduces a ratio-preserving sampler to ensure task-wise policy gradients reflect the adapted weights. Experiments on both 3-task and 9-task settings show that MT-GRPO consistently outperforms baselines in worst-task accuracy. In particular, MT-GRPO achieves 16-28% and 6% absolute improvement on worst-task performance over standard GRPO and DAPO, respectively, while maintaining competitive average accuracy. Moreover, MT-GRPO requires 50% fewer training steps to reach 50% worst-task accuracy in the 3-task setting, demonstrating substantially improved efficiency in achieving reliable performance across tasks.