This paper tackles the challenge of evaluating whether large language models perform genuine epistemic reasoning—reasoning about knowledge and partial observations in multi-agent systems—or simply rely on memorization of classic puzzles like the Muddy Children problem. The authors persuasively argue that memorization is better understood as a special case of reduction, where models map new instances to known problems. They introduce a reduction ladder with progressively modified puzzle variants to distinguish reductive from epistemic reasoning, finding that while some models succeed through reduction, all struggle when true epistemic reasoning is required. The work reframes how we interpret LLM performance on canonical reasoning benchmarks and highlights that strong accuracy on classic puzzles may mask a lack of genuine reasoning capability.

This paper addresses the high computational cost of deploying Large Language Models (LLMs) in resource-constrained environments by introducing the Performance-Efficiency Ratio (PER), a novel metric that integrates accuracy, throughput, memory, and latency via geometric mean normalization. The authors evaluate 16 open-source language models ranging from 0.5B to 72B parameters across five NLP tasks (IMDB, HellaSwag, ARC-Easy, SQuAD 2.0, and GSM8K), concluding that small models (0.5–3B parameters) consistently achieve superior PER scores compared to their larger counterparts.

TiCo tackles a critical gap in spoken dialogue models: the inability to control response duration, which is essential for time-constrained scenarios like driving assistants or emergency healthcare. Unlike text length control, speech duration depends on complex factors including phonetics, prosody, and speaking rate. The paper proposes Spoken Time Markers (STMs)—special tokens like <15.0 seconds> inserted during generation—to enable real-time temporal awareness. Using a two-stage post-training framework (self-generated supervised fine-tuning followed by reinforcement learning with verifiable rewards), TiCo equips models to estimate elapsed time and adjust content dynamically to meet target durations.

DATASHI is a new parallel corpus for Tashlhiyt, a critically under-resourced Amazigh language spoken by millions in Morocco but lacking standardized digital resources. The paper introduces 5,000 English–Tashlhiyt sentence pairs, including a 1,500-sentence subset with expert-standardized and non-standard user-generated versions, designed to benchmark orthography normalization. Using this corpus, the authors evaluate five state-of-the-art LLMs (GPT-5, Claude-Sonnet-4.5, Gemini-2.5-Pro, Mistral, Qwen3-Max) on the normalization task, finding that even the best model (Gemini-2.5-Pro) achieves only moderate accuracy (35.5% WER) and struggles with gemination and emphatic consonants.

This paper introduces BHDD, the first public benchmark dataset for handwritten Burmese digits. Myanmar script's distinctive circular letterforms—originally developed for writing on palm leaves—create recognition challenges distinct from Latin digits, with pairs like 0 and 1 differing only by whether a circle is closed. The authors release 87,561 verified images (28×28 grayscale, MNIST-compatible format) from over 150 contributors, with writer-independent train/test splits and baseline models reaching up to 99.83% accuracy.

This paper addresses cross-lingual knowledge graph fusion, where heterogeneous KGs in different languages must be unified without expensive manually-curated seed alignments. The core idea is to use Large Language Models as a universal semantic bridge by linearizing graph triplets into natural language sequences and sequentially agglomerating multiple graphs. This matters because it promises zero-shot alignment capability for low-resource languages where traditional embedding-based methods fail due to lack of training data.

Psychiatric symptom identification from social media requires expensive expert annotation and suffers from inconsistent labeling across platforms. SynSym addresses this by using GPT-4o to generate synthetic training data across four stages: symptom concept expansion, dual-style (clinical/colloquial) expression generation, clinically-grounded multi-symptom composition, and LLM-based quality filtering. The framework produces 18,254 samples covering 14 DSM-5 symptoms, enabling models to match real-data performance and generalize across diverse social media platforms.

Audio-enabled large language models promise to democratize AI access for users with disabilities or limited literacy, but voice interfaces introduce immutable paralinguistic cues—pitch, timbre, prosody—that carry demographic signals. This paper demonstrates that state-of-the-art audio LLMs systematically discriminate based on speaker voice, assigning gender-stereotyped adjectives and professions solely from acoustic features. Crucially, the authors show that voice inputs amplify bias beyond text-only baselines, with models exhibiting stronger stereotypical associations when processing speech than when processing equivalent text with gendered name cues. The study establishes a causal link via pitch manipulation experiments and surveys 1,000 users to reveal that those who would benefit most from voice accessibility are often most hesitant about the attendant privacy and discrimination risks.

SLURP-TN introduces a Spoken Language Understanding (SLU) dataset for Tunisian Arabic, a low-resource dialect. The authors translate and record six domains from the English SLURP corpus with 55 speakers across 18 geographic regions, emphasizing gender balance and code-switching phenomena. The dataset provides approximately five hours of audio across three acoustic conditions (clean, noisy, headphone) to enable robust benchmarking of ASR and SLU systems for dialectal Arabic.

MemAPO addresses a critical limitation in automatic prompt optimization (APO): existing methods frame optimization as an isolated search for task-specific prompts, preventing knowledge reuse across tasks. The paper proposes reframing APO as a continual experience accumulation process using a dual-memory mechanism—Correct-Template Memory ($\mathcal{E}_{\mathrm{CTM}}$) for successful strategies and Error-Pattern Memory ($\mathcal{E}_{\mathrm{EPM}}$) for failure modes—that enables cross-task generalization while reducing optimization costs by approximately 57% compared to strong baselines.

The paper introduces Dissimilar Span Detection (DSD), a new task aimed at explaining Semantic Textual Similarity (STS) scores by identifying specific text spans that differ in meaning between sentence pairs. To enable this research, the authors release the Span Similarity Dataset (SSD), containing 1,000 semi-automatically annotated samples validated by human annotators. They evaluate a broad range of approaches—including LIME, SHAP, proprietary LLMs, and supervised token classifiers—and find that while LLMs achieve the highest performance, the task remains challenging even for state-of-the-art models, with potential applications in paraphrase detection and fact-checking.

Large language models have historically lagged behind specialized encoder-decoder MT systems, but their superior context modeling makes them natural candidates for document-level translation. This paper tackles two key obstacles: the scarcity of high-quality document-level parallel corpora and LLM tendencies toward hallucinations and omissions. The authors propose a two-stage fine-tuning framework that first generates synthetic document-level data from summarization corpora via LLM augmentation, filters this data using sacreBLEU, COMET, and LaBSE cosine similarity, and then trains models first on sentence-level data before adapting to the filtered document corpus.

This paper distinguishes different forms of reasoning by the structural properties they demand from underlying representational systems. The core insight is that deduction requires four specific properties (operability, consistency, structural preservation, and compositionality) that cannot be secured through mere statistical scaling. This has significant implications for AI systems and cognitive science, providing a principled boundary between reasoning that can rely on associative approximations versus reasoning requiring structural guarantees.

This paper attacks the expensive problem of annotating NLP test sets by importing Active Testing (AT) from computer vision into language tasks. Given a labeling budget $B$, the goal is to select a subset $X_A$ that minimizes the estimation error $|M(X_F) - M(X_A)|$ between full and sampled test-set metrics, potentially cutting annotation costs by up to 95% while keeping prediction error under 1%. The core mechanism couples importance-weighted unbiased estimators with acquisition strategies (including a novel Agreement strategy based on attention-head disagreement) and an adaptive stopping criterion that removes the need to pre-specify the budget.

SecureBreak introduces a response-level safety dataset designed to detect harmful LLM outputs that bypass alignment mechanisms. Unlike existing benchmarks that classify prompts, this work focuses on binary classification of generated responses (safe vs. unsafe) across 3,059 samples from multiple model families including Llama, Qwen, Gemma, and Mistral. The core value proposition is providing a 'last-line defense' layer for post-generation filtering and supervisory signals to guide security re-alignment, addressing the growing threat of jailbreak attacks.

The paper addresses the brittleness of in-context learning (ICL) to example ordering, an intractable $n!$ search problem. It proposes PLR, which reframes discrete permutation search as learning a Plackett-Luce distribution that concentrates probability mass on high-performing orderings. Using Gumbel perturb-and-sort for efficient sampling, PLR optimizes task-level metrics directly without requiring finite label spaces, extending naturally to open-ended reasoning tasks like mathematical problem solving.

This paper introduces TaigiSpeech, the first intent recognition dataset for Taiwanese Hokkien—a low-resource language spoken by 65% of Taiwanese elders. With 3,000+ utterances from 21 elderly speakers across emergency and smart-home scenarios, it addresses a critical gap in speech technology for aging populations. The authors also propose keyword-based and audio-visual mining strategies to bootstrap training data from unlabeled video sources.

Symbolic regression search spaces suffer from structural redundancy: expression DAGs with $k$ internal nodes admit $\Theta(k!)$ distinct node-numberings that encode the same mathematical expression. This paper proposes IsalSR, a representation framework that computes a pruned canonical string—a complete labeled-DAG isomorphism invariant—to collapse all equivalent forms into a single canonical representation. The approach promises to reduce effective search space size by $O(k!)$ and can be integrated into any existing SR algorithm as a preprocessing step.

TIDE is a post-training early exit system for autoregressive LLMs that trains lightweight router MLPs to predict which tokens can safely exit at intermediate layers. The key idea is using cosine similarity between checkpoint hidden states and final layer outputs as a convergence signal, eliminating the need for costly model retraining. Unlike prior early-exit methods that require training from scratch or use unreliable confidence heuristics, TIDE claims to work with any HuggingFace causal LM while preserving KV cache integrity and achieving up to 8.1% throughput improvement.

Developing optimized CUDA kernels is critical for generative AI but remains challenging even for human experts. This paper introduces DRTriton, a framework that trains a 7B-parameter LLM to convert PyTorch code into efficient Triton kernels using exclusively synthetic data. The approach combines a constraint satisfaction algorithm for program generation (CSP-DAG), curriculum reinforcement learning with decoupled rewards (DRPO), and test-time search, achieving 92% speedup on KernelBench Level 2 compared to 23% for GPT-5.2.