Hybrid retrieval—fusing lexical BM25 with dense neural retrievers—is the default recommendation in modern search stacks, usually justified by aggregate benchmark gains. This paper challenges that consensus by diagnosing what it calls the "fusion tax": the frequency and severity with which fusion hurts individual queries relative to the better of its two components or to a deployable dense-only baseline. Through exhaustive experiments across seven BEIR corpora, five encoders spanning 22M–596M parameters, and multiple fusion functions, the authors show the tax is material, pervasive, and undiminished by encoder scale, and they derive practical guidance for when to fuse and which method to choose.

Time series databases store billions of points, yet answering natural-language shape queries like "plunge and then recover quickly" remains hard because large language models must bridge free-form language to raw numeric windows. TT-SAX proposes a minimal redesign of SAX that adds a second track of trend letters measured across neighboring segments, letting an LLM compile shape queries into regular expressions that can be matched efficiently at database scale. The paper is significant because it reframes symbolization design around LLM usability rather than classical distance or compression metrics.

Users increasingly rely on AI assistants for complex, multi-turn tasks, but evaluating these systems requires simulators that can both follow intricate intents and behave like real humans. This paper proposes UserLM-2.0, an 8B-parameter simulator trained with multi-turn reinforcement learning that jointly optimizes task fidelity and naturalness through a lightweight, dynamically retrained discriminator. By reaching $46.2\%$ conversation validity—over triple prior state-of-the-art—it demonstrates a practical path forward, though a sizable gap to real human conversations remains.

Text-to-speech systems are commonly judged by feeding synthetic speech into an automatic speech recognizer and measuring word error rate against the input text, a procedure that inevitably mixes ASR mistakes with true synthesis errors. This paper isolates the ASR component of that verifier's floor by running the same three recognizers on the original human recordings from LibriSpeech and FLEURS, then comparing three scoring rules: the standard ASR round-trip, a single-recognizer mask that ignores words missed in human speech, and a three-recognizer adjudication. The work matters because it shows that a large share of apparent rare-word TTS error is purely recognizer failure, and it provides principled recommendations—such as reporting per-frequency ASR baselines and avoiding the single-recognizer mask—for fairer TTS evaluation.

Most multimodal benchmarks test perception or reasoning over determinate inputs, leaving cross-modal ambiguity resolution largely unmeasured. TRIAD introduces a diagnostic benchmark of 327 text-image-audio triplets formally constructed so that only the full triplet yields a unique answer, while any two modalities leave multiple options viable. By organizing items into 86 scene groups and evaluating 14 omni-modal systems under closed and open-ended conditions, the paper shows that strong query-level accuracy masks fragile scene consistency and that removing answer options collapses models to single-digit group accuracy, reframing evaluation around tri-modal disambiguation rather than recognition.

Wearable devices generate rich physiological records, but turning that structured data into patient-readable narratives risks semantic drift that can distort clinical meaning. This work formalizes structured-to-narrative summarization for menstrual health monitoring and proposes a constrained generation plus retrieval-augmented pipeline to preserve clinical alignment across five defined semantic drift conditions. Across 4,500 generated summaries, the authors find that explicit grounding constraints improve alignment, especially uncertainty preservation, yet they come at a measurable cost in readability and increased medical terminology.

Most benchmarks for AI tutors assume cooperative, cognitively focused students and ignore the challenging behaviors—refusal, distress, disengagement—that make real tutoring difficult. TutorCaseBench addresses this gap by mining qualitative case studies from pedagogy, educational psychology, and special education, converting them into 923 grounded tutor–student dialogues with case-specific rubrics. The results show that even top LLMs fail to align with documented interventions in nearly half of open-ended cases, revealing a critical weakness in context-sensitive support for vulnerable learners.

Decision models such as JEV promise efficient zero-shot classification by predicting over predefined labels instead of generating free-form text, yet their behavior on Vietnamese tasks has been largely uncharted. This study benchmarks four specialized decision models across 13 Vietnamese classification settings spanning topic, emotion, hate-speech, sentiment, intent, and NLI tasks. The authors systematically test English, Vietnamese, and bilingual label descriptions, and they introduce three realistic linguistic perturbations—teencode, typographical noise, and diacritic removal—to probe robustness. The core takeaway is that JEV dominates overall but remains vulnerable to accent stripping, while bilingual label descriptions do not reliably help any model.

Code-switched ASR is bottlenecked by scarce real transcribed speech, while synthetic data often mismatches real acoustics and switching patterns. SyRA-CS proposes using Reptile meta-learning on synthetic CS-FLEURS speech to learn a pair-specific initialization that adapts to just one hour of labeled real target data. The paper evaluates this across four architectures and four language pairs, aiming to show that episodic meta-training extracts more transferable knowledge from synthetic data than ordinary continual fine-tuning.

Modern end-to-end ASR systems typically rely on simplistic fixed-size beam search, assuming that large search spaces are unnecessary. This paper systematically compares search complexity and efficiency for CTC and neural transducer models across three datasets, showing that dynamic score-based pruning adapts to local uncertainty and dramatically cuts the real-time factor without sacrificing accuracy. The work challenges the prevailing assumption that small fixed beams are sufficient and provides tuned decoding parameters for reproducibility.

This paper identifies a hidden cost of LLM safety alignment: when models are prompted with protected attributes such as race, they often refuse to judge the associated information at all. The authors formalize this as invisibilization using a two-urn Bayesian trust game with an exact benchmark, randomizing only the reporter's description across 6,600+ sessions. They show that some model configurations discard identical evidence far more often for Black-described reporters (46%) than white-described ones (9%), while ironically trusting the assessed Black source slightly more. The result is a probability bias that standard fairness audits miss entirely because they look only at cases where the model actually renders a judgment.

MixJudge tackles the problem of assessing vocal mixes via natural language queries, a capability that existing audio-language systems do not separately test. The authors synthesize a curriculum of 11 mixing issues at multiple severity levels and train a contrastive audio-text critic that learns to retrieve issues, discriminate opposites, and order severity. Their menu-selection experiments further show that issue evidence transfers to ranking real processing candidates against professional mixes, offering a concrete form of explicit, comparable mixing assistance.

Clinical question answering in Indonesian is hampered by a complete lack of public, domain-specific data. IndoNutriQA introduces 911 expert-authored, abstractive question–answer pairs grounded in real clinical-nutrition reports, using eight standardized questions from the Nutrition Care Process to create a controlled evaluation benchmark. Its rigorous validation protocol and public release aim to advance low-resource clinical NLP and nutrition-focused decision support.

Most cultural evaluations of LLMs reduce user identity to a single country label, ignoring that real users hold multi-layered identities. This paper introduces a composite-identity setup where a system-prompt persona names one country while a user turn signals a second identity, either explicitly by naming a country or implicitly through query language. By probing three open-weight models with WVS multiple-choice questions and 1.6 million open-ended responses, the authors show that language and country cues shift model values in different, non-interchangeable ways—a finding with direct implications for how multilingual and multicultural personalization should be engineered.

Context augmentation methods are widely used to improve large language model performance at inference time, yet it remains unclear whether reported accuracy gains reflect genuine reasoning improvements or mere statistical noise. This paper introduces a systematic sentence-level evaluation framework spanning five intervention types—from blank filler tokens to teacher-generated explanations—across seven models and three question-answering datasets. By decomposing accuracy into correct flips ($\mathrm{Corr}$) and broken predictions ($\mathrm{Brk}$) and applying a sign-randomisation test, the authors show that many gains are undirected perturbations rather than directed progress, especially for methods that add computation without new information.

The paper asks whether language models can transfer hierarchical syntactic generalizations across languages. It introduces a cross-lingual poverty-of-stimulus (CPOS) design where positive evidence for hierarchical subject-verb agreement is present only in the source language (L1), and tests transfer to an ambiguous target language (L2). The core finding is that transfer is weak and only reliably detectable when the two languages share vocabulary tokens, challenging the view that models abstract language-universal hierarchical rules independently of lexical overlap.

IQuest-Q1 is a 320-billion-parameter sparse Mixture-of-Experts foundation model (15 billion activated) optimized for agentic command-line interaction. It merges Stable SplitNorm, the Muon optimizer, and $\mu$P parameterization within a three-stage pipeline—progressive pre-training, context-extending mid-training, and multi-scaffold post-training using reinforcement learning and multi-teacher distillation—to advance open-weight CLI agent capabilities. While the paper integrates several validated training techniques and releases model weights openly, benchmark results reveal it still trails frontier proprietary systems on the hardest coding and reasoning tasks.

Procedural mistake recognition usually gets tangled up in temporal segmentation and low-level vision features. This paper introduces PRISM, a language-grounded framework that maps video steps to text, optionally structures them with semantic roles, and decodes sequences with a lightweight hidden Markov model trained only on correct executions. By assuming benchmark-provided step boundaries, it isolates the core reasoning problem of detecting deviations in step content and order through a unified edit-based taxonomy.

Multilingual web-corpus pipelines such as FineWeb2 apply one recipe across thousands of languages, leaving individual languages’ failure modes unexamined. This paper proposes treating an LLM as a one-time debugger rather than a classifier: a budgeted, tool-using agent explores the Slovak slice of FineWeb2 through a read-only API and compiles each finding into a cheap executable test validated by a native speaker. The work matters because it replaces a linear, manual audit with a reusable test suite that scans 26.5M documents in under a second without a GPU.

Do people who speak different languages organize concepts in similar ways? The authors extract word embeddings from multilingual transformers (mBERT and XLM-R-large) to compare how 196 basic concepts relate to one another across 81 languages. Their core finding is that sociocultural meanings are more aligned across languages than nature and motion meanings, a domain-sensitive pattern that recurs across encoders and concept groupings. The work connects computational linguistics to cognitive science, though its inferences about human conceptual organization remain indirect because they rely entirely on neural network representations rather than behavioral or neural data from speakers.