Small language models offer a compact basis for shell simulation without executing submitted commands, but full-output decoding requires them to predict recurring utility structure alongside content that varies across executions. We introduce template gap reconstruction (TGR), which supplies reusable output literals interspersed with sentinels to be interleaved with autoregressively generated gaps. A template selector chooses which template to be reconstructed, with the empty template retaining full-output decoding. We construct a template bank from 18.7 million outputs collected across heterogeneous execution environments and train a template selector and five reconstruction models initialized from various small language model checkpoints on a target environment excluded from bank construction. Evaluation demonstrates that TGR improves output edit similarity (O-ES) over full-output decoding (FOD) when the optimal template covers more than 25\% of the template, while performing similarly for lower-coverage outputs. Additionally, TGR outperforms FOD for shorter gap sequences, with O-ES values of 0.964 and 0.812, respectively, while maintaining similar performance for longer sequences. TGR reduces the number of tokens generated by 26.3--38.3\% across used tokenizer families. These results support using a reusable output structure as a practical design choice for small shell simulators, improving response fidelity and shortening generation targets while keeping template selection and residual prediction central to response quality.
QA-based text embeddings represent texts as vectors consisting of 1/0 answers to yes/no questions. Although these embeddings provide detailed information for interpretation and achieve high task performance, they often require humans to examine a large number of questions. Furthermore, the relationship between the number of questions and the appropriate level of question granularity has not been sufficiently studied. To address these challenges, we propose a method for constructing hierarchical questions for text embeddings. As the hierarchy deepens, questions become more specific and their number increases. During interpretation, we reduce the number of questions that humans need to examine by replacing questions at deeper levels with aligned questions at shallower levels. The question hierarchy is constructed through recursive clustering, contrastive prompting, and question alignment. Experimental results on general-purpose tasks show that our method improves interpretability while maintaining task performance. It is also found that, when only a limited number of questions are available, questions at shallower levels tend to achieve higher task performance than those at deeper levels, even with the same number of questions.
Steffen Castle, Nils Feldhus, Christopher Ebert, Leonhard Hennig, Sebastian Möller
Summarization with a fixed character length is important for tasks such as headline and article preview generation, where text must fit an exact space. Autoregressive language models (LMs) struggle with precise character-length control: training-based methods require task-specific fine-tuning, while zero-shot prompting strategies achieve only approximate compliance. We propose DELTA (Diffusion Expected Length Targeting Algorithm), a training-free inference-time method for masked diffusion LMs that estimates expected character length at each denoising step and dynamically inserts or removes masked positions to steer toward a target length. DELTA requires no fine-tuning and is designed for off-the-shelf masked diffusion LMs. Across three summarization benchmarks, it reduces character-length errors by an order of magnitude compared to baselines while maintaining competitive summary quality in both automatic and human evaluation.
Bo Wang, Yuxuan Zhang, Yueqin Hu, Hanchao Hou, Kaiping Peng, Shiguang Ni
Psychological scale refinement traditionally relies on response-based methods such as factor analysis, item response theory, and network psychometrics to uncover latent constructs and optimize item composition. Although rigorous, these approaches require large samples and may be constrained by data availability and cross-cultural comparability. Recent advances in natural language processing suggest that the wording of scale items contains construct-relevant semantic organization, offering a complementary response-free perspective. We introduce a topic-modeling framework that operationalizes item-level semantic structure for systematic scale simplification. Items are encoded using contextual sentence embeddings and grouped through density-based clustering to identify semantic topics without predefining their number. Within each cluster, class-based term weighting identifies representative keywords that summarize each topic and support an interpretable semantic pre-structure of the item pool. Semantically related clusters are hierarchically merged, and representative items are selected using structure-aware membership criteria within an integrated reduction pipeline. We evaluated the framework on three widely used instruments: DASS, IPIP, and EPOCH, covering negative affect, personality traits, and well-being. Results showed that the identified semantic topics often formed factor-like groupings aligned with established constructs. Representative items selected through semantic membership probabilities reduced item counts by 60.5% on average while maintaining acceptable psychometric properties and preserving factor correlation patterns. Our framework formalizes semantic topic structure as an inspectable front-end for scale construction and simplification. To facilitate adoption, we provide a visualization-supported tool enabling one-click semantic analysis and structured simplification.
Streaming speech recognition is increasingly consumed by machines, such as translators and voice agents, that act on each token as it appears. A token that is later revised costs wasted work or a wrong action. In this work, we present Confucius4-R2T2 (short for Real Real-Time Transcription), a stable-prefix streaming ASR model that never modifies its earlier output while maintaining high accuracy and low latency. R2T2 is obtained by post-training Qwen3-ASR-1.7B on 1,500 hours of Chinese and English audio while freezing its pre-trained audio encoder and projector. The adaptation is based on the concept of Longest Stable Prefix (LSP), a new notion that we introduce in this work. We prove that LSP is monotone under the prefix order, consistent with the full-audio output, and maximal among labels that equal a whole earlier output. Learning from LSP labels converts a full-context model into a stable-prefix streaming one without any architectural change. Extensive evaluations show that R2T2 brings streaming recognition close to the base model's full-context quality while delivering stable text with low emission lag. Its gains extend beyond the post-training languages, demonstrating a practical path from full-context ASR to stable streaming. We open-source the model weights and code at https://huggingface.co/netease-youdao/Confucius4-R2T2 and https://github.com/netease-youdao/Confucius4-R2T2.
Despite their remarkable capabilities, large foundation models (LFMs) remain susceptible to adversarial manipulation. Current safety intervention methods primarily rely on the \textit{locality hypothesis}, rigidly suppressing isolated neurons or features. However, harmful semantics manifest as distributed, cross-layer circuits, rendering localized ablations both brittle against complex jailbreaks and detrimental to general utility. To bridge this gap, we propose TraceRouter++, a routing-level safety intervention framework that explicitly traces safety-relevant causal routing pathways activated under harmful inputs and realigns their unsafe activation states. Our framework operates in a \textit{Discover-Trace-Realign} paradigm: it first pinpoints a sensitive onset layer via attention divergence and disentangles dense activations into interpretable routing features using sparse autoencoders (SAEs). Subsequently, it maps these features to downstream causal circuits via feature influence scores (FIS). Crucially, departing from rigid suppression paradigms, TraceRouter++ introduces a constrained optimal transport (OT)-based soft routing realignment. By formulating safety intervention as a tractable coordinate-wise marginal optimal transport approximation, our method redirects unsafe activation marginals toward safe-reference marginals within causally identified routing subspaces while preserving non-selected coordinates at the intervention point. This mechanism safely redirects malicious semantic flows while leaving orthogonal computational routes intact. Extensive experiments across representative large language models (LLMs), multimodal large language models (MLLMs), and diffusion models (DMs) show that TraceRouter++ outperforms evaluated baselines, improving adversarial robustness while maintaining a favorable safety-utility trade-off. WARNING: This paper contains unsafe model responses.
Jingcheng Niu, Xingdi Yuan, Tong Wang, Hamidreza Saghir, Amir H. Abdi
We observe a novel phenomenon, contextual entrainment, across a wide range of language models (LMs) and prompt settings, providing a new mechanistic perspective on how LMs become distracted by ``irrelevant'' contextual information in the input prompt. Specifically, LMs assign significantly higher logits (or probabilities) to any tokens that have previously appeared in the context prompt, even for random tokens. This suggests that contextual entrainment is a mechanistic phenomenon, occurring independently of the relevance or semantic relation of the tokens to the question or the rest of the sentence. We find statistically significant evidence that the magnitude of contextual entrainment is influenced by semantic factors. Counterfactual prompts have a greater effect compared to factual ones, suggesting that while contextual entrainment is a mechanistic phenomenon, it is modulated by semantic factors. We hypothesise that there is a circuit of attention heads -- the entrainment heads -- that corresponds to the contextual entrainment phenomenon. Using a novel entrainment head discovery method based on differentiable masking, we identify these heads across various settings. When we ``turn off'' these heads, i.e., set their outputs to zero, the effect of contextual entrainment is significantly attenuated, causing the model to generate output that capitulates to what it would produce if no distracting context were provided. Our discovery of contextual entrainment, along with our investigation into LM distraction via the entrainment heads, marks a key step towards the mechanistic analysis and mitigation of the distraction problem.
We propose Online Latent prediction with Invariant Views and rEconstruction (OLIVE), a self-supervised speech representation learning framework that jointly optimizes analysis and synthesis objectives. OLIVE couples masked latent prediction and waveform reconstruction through a functional separation within a single encoder: reconstruction constrains early encoder features to retain signal-level information, while masked latent prediction shapes later contextual representations for robust downstream performance, with waveform view augmentations defining the invariances they learn. To the best of our knowledge, OLIVE is the first speech SSL model that directly reconstructs waveforms, retains its vocoder after pre-training, and remains competitive on discriminative tasks. Adding reconstruction largely preserves the speech recognition performance of masked latent prediction alone, while improving waveform reconstruction and speaker tasks. Adding waveform mixup and gain augmentations provides a controllable way to shift the representations further toward speaker, generation, and reconstruction tasks, while remaining competitive on recognition and semantic tasks.
LLM-as-judge evaluation has become standard practice for open-ended model assessment; however, judges exhibit systematic biases that cannot be averaged out by increasing the number of scenarios or generations. These biases are often similar in magnitude to the model differences that benchmarks are designed to detect, resulting in unreliable rankings when single-judge evaluations are used. We introduce a variance decomposition that partitions benchmark score variance into scenario, generation, judge, and residual components. Based on this analysis, CyclicJudge, a round-robin assignment of judges to scenarios, is demonstrated to be the optimal strategy for a fixed judge panel and judge-call budget: it cancels the average additive judge offsets while matching the cost of single-judge evaluation. Empirical results on MT-Bench and MindEval validate the effectiveness of CyclicJudge as predicted, across both general-purpose and domain-specific evaluation settings, reducing benchmark score variance and stabilizing the resulting model ranking.
Ziyi Zhu, Olivier Tieleman, Caitlin A. Stamatis, Luka Smyth, Thomas D. Hull, Daniel R. Cahn, Jinghong Chen, Matteo Malgaroli
Realistic user simulation is crucial for training and evaluating multi-turn dialogue systems, yet creating simulators that accurately replicate human behavior remains a significant challenge. An effective simulator must expose the failure modes of the systems under evaluation. This work introduces Direct Iterative Adversarial Learning (DIAL), an adversarial framework that iteratively enhances user simulator realism through a competitive dynamic between a generator (user simulator) and a discriminator. When applied to mental health support, a domain characterized by diverse failure types and a critical dependence on realistic user behavior for failure detection, DIAL restores lexical diversity diminished by supervised fine-tuning and drastically reduces discriminator accuracy. The resulting simulator exhibits a strong correlation between simulated and real failure occurrence rates while maintaining low distributional divergence of failure modes. These findings indicate that DIAL is a promising method for developing realistic user simulators in multi-turn dialogue, facilitating reliable and cost-effective system evaluation prior to deployment.
Balkrishna Giri, Md Toufique Hasan, Jussi Rasku, Muhammad Waseem, Pekka Abrahamsson
Retrieval-Augmented Generation (RAG) grounds Large Language Model (LLM) outputs in external knowledge, but RAG systems usually trust whatever they retrieve, creating a Security-Reliability Gap: high semantic relevance does not guarantee factual truth. Adversaries exploit this through knowledge poisoning, inserting malicious documents to cause targeted misinformation. We propose an Evaluation Agent, middleware that combines Natural Language Inference (NLI) factual verification, a five-signal poison detector with relevance-weighted aggregation, and a Trust Index T = 0.4 F + 0.35 C + 0.25 (1 - P ) with a non-linear dampener for high-contamination contexts. On TruthfulQA with Llama 3.3 70B, the agent reaches 91% accuracy and 100% precision, with 100% recall on instruction injection, while in-place edits, such as entity swaps, remain hard to detect. Across three LLMs the Trust Index stays discriminative, with a Receiver Operating Characteristic Area Under the Curve (ROC-AUC) of 0.73 to 0.81; generation style matters more than model size, and per-LLM threshold calibration restores baseline competitive accuracy, whereas a weaker FEVER result shows that cross-dataset generalization requires domain-specific calibration. In a software-engineering use case, a secure-coding assistant over guidance from the Open Worldwide Application Security Project (OWASP) Top 10 and the Common Weakness Enumeration (CWE), the agent reliably blocks instruction injection of unsafe advice (F1 92%), while contradiction and subtle semantic weakening remain hard. Throughout, the agent measures detection of poisoned context before generation, not whether the LLM adopts the injected misinformation. We release the proposed approach, attack generator, and experimental artifacts at the link: https://github.com/GPT-Laboratory/TrustworthyRAG.
Fast rollout generation for long-context reinforcement learning with verifiable rewards (RLVR) requires reducing KV-cache traffic without changing the response groups used for policy optimization. In our profiled RLVR workloads, attention accounts for 72--76\% of autoregressive decoding kernel time. Sparse attention reduces these reads, but sampling directly from a sparse policy changes the response groups used for learning. For group-relative optimization, we show that even exact response-wise reweighting does not generally recover the expected dense-policy update because the advantage of each response depends on the rewards of other group members. We propose Self-Sparse Speculative Rollouts (S$^3$R). It uses sparse attention to propose tokens, then generates responses through full-attention verification and exact speculative sampling. Drafting and verification share the same policy parameter, eliminating the need to train or update a separate drafter. The resulting response groups follow the dense-policy distribution. Under S$^3$R, we characterize speedup conditions and the costs of drafting and verification. We found that dense verification still requires the complete historical KV cache. A KV-aware scheduler therefore limits KV-cache repeated offloading and restoration while maintaining rollout parallelism. % OR thrashing In a matched ablation with an aggressive sparse drafting budget, S$^3$R maintains stable training while direct sparse sampling with training-side correction becomes unstable. On long-context mathematical reasoning, S$^3$R achieves $1.39\times~1.35\times$ the end-to-end training throughput of autoregressive rollouts on Qwen3 model, with comparable task performance.
Historical archives are invaluable for multidisciplinary research but remain inaccessible globally due to severe language barriers. While manual translation is prohibitively expensive and time-consuming, simply applying existing machine translation models is inadequate due to unique linguistic nuances and the temporal bias inherent in scarce human-labeled corpora. To address these challenges, we propose HERIT, a systematic domain-adaptation pipeline that leverages retrieval-augmented generation (RAG) to generate high-quality pseudo-labeled data from abundant Hanja-Korean corpora. This approach substantially expands the training dataset, mitigating both data scarcity and temporal overfitting across different historical periods. Extensive automated and human evaluations demonstrate that HERIT consistently outperforms baselines. Finally, we translate previously untranslated portions of the archives using our model, thereby enabling global access to these resources for researchers.
Ivo Petrov, Jasper Dekoninck, Dimitar I. Dimitrov, Martin Vechev
Large language models (LLMs) have become capable mathematical problem-solvers, often producing correct proofs for challenging problems. However, correctness alone is not sufficient: mathematical proofs should also be clear, insightful, contain transferable knowledge, and be generally useful to the reader. While proof quality is subjective and depends on the reader and context, many of its components can be made concrete. In this work, we identify such components and introduce ProofRank, a benchmark curated from challenging competitions that measures several scalable proxies of proof quality: (i) conciseness - quantifying whether proofs avoid unnecessary steps; (ii) computational ease - the extent to which a proof relies on tedious calculations; (iii) cognitive simplicity - how accessible the used proof techniques are; (iv) diversity - how varied a model's proofs for a single problem are; and (v) adaptivity - how well a model can follow a specified technique. We find proof quality varies substantially across models along dimensions that correctness largely fails to capture. Every proof-level proxy can be steered through prompting, but often at a moderate correctness cost. ProofRank thus captures important aspects of mathematical proofs beyond correctness, and indicates the need for models that produce proofs which are not only correct but also valuable.
Large reasoning models achieve strong performance on challenging tasks, but their long chain-of-thought often incurs substantial inference costs by continuing to deliberate after sufficient information has been obtained. Existing reasoning compression methods primarily operate on reasoning trajectories, paying less attention to the local decision of whether further reasoning is necessary given the current state. We propose LoCo, a local preference learning method for reasoning compression. LoCo detects candidate overthinking states from the model's own traces, freezes the corresponding reasoning prefixes, regenerates shorter verified continuations from the same states, and trains on the resulting state-conditioned preference pairs with DPO. This formulation teaches the model when to continue reasoning and when to conclude, rather than simply encouraging shorter trajectories overall. We construct training data entirely from self-generated mathematical traces and evaluate LoCo on five mathematical and three code generation benchmarks. LoCo reduces completion length by 41.7\%--45.4\% on mathematics and 23.4\%--36.3\% on code, while maintaining or improving accuracy and achieving the best AUC across all code benchmarks. LoCo consistently improves the accuracy-efficiency trade-off on mathematics and transfers effectively to code generation despite math-only training, demonstrating the cross-domain generality of local preference learning for reasoning compression. The code and model will be publicly released at https://github.com/MuyaoYuan/LoCo.
Large language models (LLMs) face significant challenges in processing long contexts due to the linear growth of the key-value (KV) cache and quadratic complexity of self-attention. Existing approaches address these bottlenecks separately: Multi-head Latent Attention (MLA) reduces the KV cache by projecting tokens into a low-dimensional latent space, while sparse attention reduces computation. However, sparse methods cannot operate natively on MLA's compressed latent structure, missing opportunities for joint optimization. In this paper, we propose Latent-Condensed Attention (LCA), which directly condenses context within MLA's latent space, where the representation is disentangled into semantic latent vectors and positional keys. LCA separately aggregates semantic vectors via query-aware pooling and preserves positional keys via anchor selection. This approach jointly reduces both computational cost and KV cache without adding parameters. Beyond MLA, LCA's design is architecture-agnostic and readily extends to other attention mechanisms such as GQA. Theoretically, we prove a length-independent error bound. Experiments show LCA achieves up to 2.5$\times$ prefilling speedup and 90% KV cache reduction at 128K context while maintaining competitive performance.
Siri Expressive Voices synthesize rich, configurable speech in real time and entirely on device, powered by AFM 3 Core Advanced, Apple's most powerful on-device foundation model. This work presents the memory-efficient audio synthesis architecture behind that capability: a detokenizer that converts the semantic audio tokens emitted by the foundation model into high-fidelity audio within the tight compute and memory budget of the Scalable Matrix Extension (SME). We convert semantic audio tokens to a residual vector quantization (RVQ) representation with a three-component design, a streaming encoder, a temporal decoder, and a depth decoder, that systematically decouples temporal and depth processing. A single reusable depth decoder with Diffusion Transformer (DiT)-style stage conditioning generates all RVQ levels autoregressively, replacing the dedicated per-level decoders of prior multi-decoder architectures, while causal sliding window attention with fixed-window key-value caching yields constant memory complexity independent of sequence length. Deployed on the AMX, the detokenizer sustains roughly 10 ms per generation step, about 16x faster than real time, with a peak runtime memory of only 21 MB and 329 MB of on-device assets, enabling continuous streaming synthesis of 20-320 seconds of audio. This constant, small footprint replaces the linear and quadratic memory scaling of conventional transformer- and GAN-based approaches. Ablation studies validate the key architectural components, and audio quality assessment confirms that the architecture maintains synthesis fidelity while achieving efficiency gains over existing methods. Operating at a 1-billion-parameter activation size within AFM 3 Core Advanced, it improves Mean Opinion Score by +0.28 overall (4.15 vs. 3.87) and by +0.42 on conversational speech (4.24 vs. 3.82) over the prior on-device text-to-speech system.
Large language model (LLM) coding agents increasingly operate with shell access, allowing them to write scripts, move files, alter permissions, and modify system configurations. However, an agent may achieve a user's goal while causing unintended harm, such as deleting unrelated files, weakening access controls, or corrupting data, in response to benign inputs without explicit jailbreaking. This failure mode is particularly important because apparent task success may reassure users that an agent behaved correctly, causing harmful side effects to go unnoticed. We call this harmful success: the agent successfully completes the intended benign task while simultaneously causing unintended harm. Despite its importance, harmful success is difficult to measure because conventional task-success metrics overlook harmful behaviors, while existing safety evaluations are often designed solely to assess harmful outcomes. We introduce VeriHarm, an automatic framework for constructing programmatically verifiable benchmarks that directly measure harmful success. VeriHarm minimally perturbs coding tasks to elicit unintended harmful behavior while maintaining the benignity of the task and preserving the original task objective. It then synthesizes executable harm tests that, together with the original task-correctness tests, evaluate both task accuracy and harmful outcomes. Frontier agents, including GPT-5.6-Sol and Claude-Opus-5, achieve an average task accuracy of 95% on our curated VeriHarm-Bench, yet still struggle to complete tasks safely: their safe success rate, the fraction of tasks completed successfully without triggering the designated harm, averages only 37%. These results suggest that successful task completion can frequently coexist with unintended harm, and that task-success metrics alone may substantially understate the potential safety risks of coding agents.
Jun Liu, Zhenglun Kong, Changdi Yang, Fan Yang, Tianqi Li, Peiyan Dong, Joannah Nanjekye, Hao Tang 等 15 位作者
Large language model (LLM) agents are increasingly organized into multi-agent systems for complex reasoning and decision-making tasks. However, scaling such systems can substantially increase inference cost, especially when multiple agents repeatedly process overlapping or role-irrelevant context. Many existing multi-agent frameworks focus on orchestration and role specialization, but still rely on full-context sharing or static context routing, leading to redundant token consumption, increased runtime, and inefficient memory utilization. We present \emph{Role-Conditioned Context Routing (RCR)}, an execution-time layer that allocates role-relevant shared-memory items to agents under token budgets. Given a structured shared memory and a set of role-specialized agents, RCR selectively routes task-relevant memory items to each agent according to its role, optional task-stage metadata, and token budget. RCR integrates role-conditioned relevance scoring, token-budgeted memory selection, and iterative memory updates to reduce redundant context exposure during multi-agent execution. Experiments on multi-hop QA, embodied reasoning (ALFWorld), and web-based decision-making (WebShop) show that RCR reduces token usage by 35\% on average and runtime by 38\% on average relative to Full-Context, while maintaining or improving task performance.
Dayam Nadeem, Safdar Mustafa, Ameen Ur Rehman, Shahab Saquib Sohail, Mohammad Nadeem, Usman Naseem
Large language models are increasingly consulted for religious knowledge, including the recall of specific Hadith (recorded sayings and actions of the Prophet Muhammad, peace be upon him). Unlike open-ended generation, Hadith recall is a high-stakes, verifiable task: a reference (collection, chapter, and Hadith number) has exactly one correct text, and any deviation from it constitutes a fabrication that can be mistaken for authentic scripture. This study introduces a bilingual (English/Arabic) benchmark that probes five open-weight, locally-deployable language models (Mistral-7B, Gemma3-4B, Jais-7B, ALLaM-7B, and Phi4-mini-3.8B) across all six canonical Sunni Hadith collections (Sahih al-Bukhari, Sahih Muslim, Sunan Abi Da'ud, Jami' al-Tirmidhi, Sunan an-Nasa'i, Sunan Ibn Majah), sampling 500 Hadith per collection (3,000 per model per language, 30,000 responses total). Each model is given only the bibliographic reference and asked to reproduce the exact Hadith text, with no retrieval augmentation, and is scored on lexical fidelity (BLEU) and semantic fidelity (multilingual cosine similarity), with refusals flagged separately. Across all five models and both languages, 0% of 30,000 responses achieve both high lexical and high semantic fidelity, i.e., an exact, verifiable match to the source text. Arabic responses fare uniformly worse than English. These findings suggest that current open-weight LLMs, used without retrieval grounding, are unreliable and potentially dangerous sources for exact religious-text recall, and that deployment for Islamic scholarship should mandate retrieval-augmented generation over a verified corpus rather than parametric memory. The complete code, prompts, evaluation pipeline, and scored outputs are publicly available at https://github.com/dayam8696/Hadith-Hallucination-Benchmark.
Swapnil Gandhi, Siva Kumar Sastry Hari, William J. Dally, Christos Kozyrakis
Recent methods expose intra-request parallelism in LLM outputs, allowing independent branches to decode concurrently. Existing serving systems execute these branches eagerly or under fixed caps. We show that both are brittle: eager admission inflates the shared decode step, degrading co-batched requests in serial stages, while conservative fixed caps forgo the throughput that motivated exposing branches in the first place. We call the excess step latency caused by admitted branches the \emph{branch externality} and show that the safe width depends on batch composition, context lengths, and accumulated slack, all of which change continuously over a workload trace. We introduce PACE, a per-step admission controller that treats extra branches as opportunistic width, admitted only when the predicted branch externality fits within the batch's current slack budget. Per-step regulation is practical because branch-level scheduling decouples compute from memory: branches share the request's prefix KV, so expanding or contracting width requires no memory reclamation. On Qwen3-32B, PACE improves goodput by $1.77\times$ over serving without branch parallelism and by $1.48\times$ over eager admission while maintaining 99\% SLO attainment. Comparable gains ($1.72$--$2.15\times$) hold on five additional dense and mixture-of-experts models with up to 671B parameters. Code is available at https://github.com/Swapnil-Gandhi/PACE
This work presents an evaluation of large language models (LLMs) for English to dialectal Arabic machine translation on the MADAR dataset. We evaluate both translation directions (English to Arabic and vice-versa) on 16 Arabic dialects. Our experiments cover a diverse set of models, including specialized Arabic models (Jais, Nile), multilingual models (Gemma, Command-R, Mistral, Aya), and commercial APIs (GPT-4.1). We employ multiple evaluation metrics: BLEU, CHRF, COMET (both reference-based and reference-less variants) and GEMBA (LLM-as-a-judge), as well as a small-scale manual evaluation, to assess translation quality. We discuss the challenges of automatic MT evaluation, especially in the context of Arabic dialects. We also evaluate the ability of LLMs to classify the dialect used in a text. The study offers insights into the capabilities and limitations of current LLMs for dialectal Arabic machine translation, particularly highlighting the difficulty of handling dialectal diversity, although the results may be influenced by possible training data contamination, which is always a concern with LLMs.
Backpropagation is the standard method for computing the gradients that train neural networks, but it is usually presented in a compact vector form that skips the intermediate steps. This tutorial derives backpropagation step by step for a fully connected feedforward network with sigmoid activations and a squared-error loss. Using a running example with three layers, we first differentiate the error with respect to a single weight by following its path through the network with the chain rule. We then show how summing over all paths leads to the error terms $δ$ and to the four general backpropagation equations in matrix form. Every term is annotated with its dimensions, so each equation can be checked for consistency. The derivation is then extended to $m$ training examples, where the weight gradients become a single matrix product and the bias term corresponds to broadcasting. It is also extended to mini-batch gradient descent, including why the gradient of a mini-batch estimates the full gradient. Batch and mini-batch training are both given as algorithms. Finally, each equation is linked to the corresponding line of a NumPy implementation trained on the AHDBase dataset of handwritten Arabic digits, and the complete code is included.
FOLLIE (for First-Order Logic for Language Inference and Entailment for French) connects French sentences, compositional first-order representations and automated natural language inference (NLI). The first French NLI-ready, NL-to-FOL resource, it contains 19,983 inference records and 36,987 non-empty sentence-slot--formula mappings from six heterogeneous NLI sources, retaining French lexical material and source problem structure. Its silver annotations are produced by a neural--symbolic pipeline, with 111 separately identified LLM-assisted additions. The construction combines morphosyntactic analysis, probability-conditioned categorial multitagging, grammatical proof search, discourse representation and first-order semantic construction. Stage-wise experiments show substantial gains in derivation coverage; analyses of conjunction distributions, logical-operator incidence and scope alternation characterise the resulting resource. We then develop an automatic inference engine whose input-conditioned background theories address lexical relations, quantification, quantities, event structure and temporal dependencies. Prover9 supplies proofs for entailment and contradiction, while Mace4 supports bounded consistency checks. Three-way accuracies are 65.37% on all 335 FraCaS-FR problems, 70.33% on 300 GQNLI-FR problems and 71.35% on 9,399 eligible SICK-FR records pooled across splits. Agreement with reference labels among committed decisions is 93.53%, 100.00% and 98.66%, at commitment coverage of 41.49%, 37.00% and 15.05%, respectively. Worked constructions, axiom schemas and proof analyses explain these outcomes and identify the contributions of scope, reference, lexical knowledge and numerical structure. FOLLIE provides a French experimental basis for studying the representations and assumptions that connect sentence interpretation to deductive inference.
While language models (LMs) excel in natural language tasks, their ability to learn impossible languages--those that violate the structural, cognitive, or logical constraints that make human languages learnable and functional--remains debatable. This controversy is further fueled by a divide between linguistic theorists and empirical researchers working with small-scale LMs. To address this controversy, we introduce semantic coherence. This concept measures how effectively semantic information is expressed in language structures and examines the learning capabilities of LMs across multiple scales. By fine-tuning billion-parameter LMs on impossible language samples, we demonstrate that larger LMs achieve similar performance on impossible and possible languages, whereas million-parameter LMs exhibit significant degradation when trained on impossible languages. This performance gap widens as semantic complexity increases. We further show that scaling up the LM size fundamentally shifts the learning focus from grammatical rules to semantic understanding, with billion-scale LMs exhibiting a threefold reduction in sensitivity to low-level linguistic perturbations while maintaining stable performance under high-level semantic perturbations.
Khandokar Md. Rahat Hossain, Souvik Das, Tahsina Hashem, Mohammed Eunus Ali, Tanzima Hashem
Accurate house price prediction is essential for real estate buyers, sellers, agents, and financial institutions. Traditional methods rely primarily on structured data, e.g., number of bedrooms, bathrooms, parkings, but they overlook rich contextual information embedded in textual descriptions and property photos. Previous deep learning approaches for house price prediction have fused multimodal data but they fall short to extract structured spatio-temporal and contextual features from texts and semantic visual information from images. In this paper, we introduce a novel multimodal framework that addresses these limitations by systematically integrating structured attributes, textual descriptions, and visual images of properties to achieve higher prediction accuracy while maintaining interpretability. We use a large language model to derive domain-specific spatio-temporal and contextual features from text, and an unsupervised vision-language approach to automatically group photos into meaningful classes like indoor, outdoor, mixed, or floorplan, that are commonly present in real-world house listings. These categories are created without manual labeling by aligning image clusters with carefully designed pseudo text prompts. We fuse all features into a unified representation and use an optimized regressor to predict prices. Experiments on a large real-world property dataset demonstrate that our method outperforms traditional methods and existing multimodal baselines. Ablation studies confirm that the novel textual feature extraction and categorized visual embeddings substantially contribute to more accurate and reliable price predictions.
While online medical consultation broadens access, unprompted patient disclosures and topic shifts frequently disrupt clinician inquiry, leaving clinical states progressively revealed. We characterize these dynamics using both existing evaluation metrics and task-specific interaction metrics designed in this work. Through an empirical study across 16 medical dialogue datasets, we show that authentic online interactions exhibit substantially lower contextual continuity and higher unpredictability than offline and simulated benchmarks. We propose PRISM (Policy Reasoning with Inferred States for Medicine), a framework inspired by partially observable Markov decision processes that decouples latent state estimation from action selection to address drifting dialogue trajectories. Specifically, PRISM factorizes clinical triage into belief estimation, policy optimization, and response realization. By maintaining an evolving posterior belief over consultation states and conditioning action selection on the maximum a posteriori state, PRISM advances clinical goals while accommodating patient-initiated shifts. Extensive evaluations show that PRISM substantially outperforms frontier LLMs and clinical agents, achieving a 22.40 percentage point gain in clinical decision accuracy alongside decisive improvements in response generation, surpassing the strongest baseline by 5.51 points in BERTScore_F1. Furthermore, ablation studies show that providing oracle consultation states consistently boosts response quality across all evaluated baselines by guiding their action reasoning, yielding average gains of 5.58 points in BERTScore_F1.
Attention kernels on Tensor Processing Units (TPUs) often underutilize their matrix units (MXUs) because of repeated matrix operand loads, vector-intensive computation, and dependencies that leave MXUs idle. We present FlyWheel, a set of fast TPU kernels for both softmax and linear attention. FlyWheel reformulates softmax and triangular inversion to reduce vector work, while coordinating data layouts to improve matrix operand reuse. It then overlaps independent computation to hide execution latency. We apply these techniques to both softmax attention and Gated DeltaNet on TPU v6e and v7x. FlyWheel speeds up softmax attention by up to $4.14\times$ over the state-of-the-art RPA kernel and $1.80\times$ over Splash Attention, reaching 81% of peak BF16 throughput on v6e. It also speeds up Gated DeltaNet prefill by up to $3.61\times$. Integrated into vLLM serving Qwen3.8-27B, FlyWheel improves prefill throughput by up to $1.56\times$ while maintaining comparable accuracy.
Voice conversion (VC) models can be used to anonymize speech data while preserving naturalness. Existing models, however, are mainly optimized for English and other high-resource languages, and they fail to generalize well to Arabic due to its distinct phonology. In this work, we comprehensively evaluate existing zero-shot VC systems for Arabic speech, identifying limitations in pronunciation preservation. We then train an Arabic-centric model based on EZ-VC on diverse Arabic speech, and demonstrate strong performance in intelligibility, naturalness, and pronunciation preservation, while maintaining high speaker dissimilarity. Through subjective evaluation, we further demonstrate its robustness to code-switching and cross-dialectal conversion, as well as its generalization to Arabic dialects with limited representation in the training data. Our findings highlight the limitations of zero-shot VC and the importance of target-language adaptation for robust voice anonymization.
Homebrewed SLMs: Autoregressive and Masked Diffusion Language Models Built at Home for Portuguese
Pedro Carvalho Brom
AI 速览：一篇教程式论文，展示如何在单块消费级 GPU 上从零训练含分词器的葡萄牙语小型语言模型，并用预注册协议比较自回归与掩码扩散两种生成方式在专业法律任务上的表现。
Building a language model of one's own no longer requires a data centre. This tutorial trains one from scratch, tokenizer included, on one consumer GPU, and shows how to tell whether it is good for the tasks of a specialist domain. It proposes no new method. It releases a tested Python package and two notebooks in Brazilian Portuguese with autoregressive and masked diffusion generation on one network, a preregistered comparison protocol, legal tasks built from text alone with a Kneser-Ney floor and public yardsticks, reference numbers for models of 11 to 12M parameters, and a catalogue of defects that only measurement revealed. On Brazilian prose and federal laws, at minutes per run, the autoregressive mode peaks at 64 epochs over the same text and then memorises it, while masked diffusion improves up to 256. At their best points on test, the diffusion bound exceeds the autoregressive likelihood by 0.06 bits per byte, within the preregistered minimum effect. On held-out laws and on acts published after training, the model ties a 6-gram floor in likelihood on the same windows, and it loses to the floor at continuing those acts. It beats the floor at detecting tampered text, and fills cross-references more often than the floor and public models of 160M and 630M parameters, only in decrees, a genre absent from training, where those scorers stay near or below chance, and in filling the blank only when each group weighs the same, since per item it ties the 630M model. On laws, the public models, trained with about 5x10^2 to 10^5 times its compute, lead in tampered text and continuation. The blind human reading is preregistered and not part of this version.