All papers

This paper presents CacheFuse, a training-free technique enabling heterogeneous language models to communicate directly by sharing key-value cache states.

This paper proposes DeepThink, a conceptual architecture that decouples knowledge capacity, activation width, and recurrent depth in language models.

This paper introduces Lüshi-1840, a pre-modern Chinese language model trained on historical texts to evaluate knowledge cutoffs and hindsight bias.

This paper analyzes prediction-derived error updates and test-time gradient objectives in trained test-time-training layers.

This paper demonstrates that deep speech representations track word timing along low-dimensional circular trajectories across multiple model architectures.

This paper proposes RIFT, a method using electrical resistance analogies to trace token influence through transformer attention graphs.

This paper introduces LoongSage, an asynchronous reinforcement learning framework designed to run multi-turn agent tasks with reduced system overhead.

This paper proposes verbal regularization prompts for on-policy self-distillation to prevent teacher models from taking shortcut reasoning paths.

This paper introduces HexaBee-4B, a translation model trained across six languages using joint multilingual answering and bidirectional supervision.

This paper presents PersonaFactors, a framework that distills reusable preference patterns across tasks for dynamic prompt personalization in language models.

This paper introduces an interactive retrieval-augmented framework that combines knowledge graphs and text retrieval to explore digital cultural heritage archives.

This paper introduces a multi-agent sandbox benchmark to evaluate language models on state inference, forecasting, and intervention prediction.

This paper presents a case study of an interpreter co-creating a local computer-assisted interpreting tool using large language models.

This paper introduces UNIGRAPHE, a deterministic method to assemble clinical text into queryable knowledge graphs conforming to SNOMED CT grammar.

This position paper argues that large language models perform language as social practices rather than building distributed representations of symbols.

This paper examines neuron relevance characterization methods based on magnitude and distribution support for selecting instruction fine-tuning data.

This paper investigates gradient spectral properties during reinforcement learning with verifiable rewards to design an adaptive-rank optimizer for reasoning models.

This paper presents a self-supervised recursive hypernetwork framework using parse trees to learn syntactically informed compositional sentence representations.

This paper introduces a Chinese multi-task evaluation benchmark covering six recruitment application areas for language models.

This paper introduces an environment benchmark and exploration framework to help graphical user interface agents learn unfamiliar interaction dynamics.

This paper constructs joint token probability laws from physics models to improve speculative verification in parallel language model decoding.

This paper proposes a cumulative value parameterization method for reinforcement learning to improve credit assignment in multi-step language reasoning.

This paper uses Bell's theorem bounds to analyze capacity limits and task interference in multi-task shared representations.

This paper presents a language model framework to extract event representations from autobiographical narratives into geometric state spaces.

This paper uses survival analysis on Reddit text and activity to predict when users will seek mental health support.

This paper studies cross-entropy surprisal metrics, showing that predictability rankings of texts remain consistent across multiple small language models.

This paper introduces a benchmark and rule-based metric for generating and evaluating audio descriptions in university slide lectures.

This paper evaluates encoder fine-tuning and synthetic data augmentation for fallacy detection and argument classification on Reddit discussions.

This paper proposes a mathematical framework using cellular sheaves to isolate phrasing-invariant content representations in transformer language models.

The paper proposes an agent framework that discovers lightweight proxy evaluators to bypass costly benchmark evaluations during autonomous engineering discovery.

The paper proposes combining teacher-forced embeddings with soft tokens to improve supervised fine-tuning on reasoning datasets without changing inference.

The paper introduces an attention conversion method enabling multi-head latent attention on modern architectures with reduced memory footprints.

The paper formulates online language model persona personalization as a dueling bandit problem using sequential pairwise user preferences.

The paper introduces a 174-hour speech benchmark and evaluates baseline models on Ewe automatic speech recognition.

The paper introduces a framework combining guidelines and agentic linking to ground medical concepts across diverse text types and ontologies.

The paper introduces a lightweight guidance model to estimate parallel decoding inconsistency and accelerate diffusion language models.

The paper evaluates five language models on extracting complete interval networks of relative temporal events from narratives.

The paper proposes a knowledge distillation method that transfers text language model capabilities to speech-to-speech models using internal text interfaces.

The paper studies compute allocation across multi-agent coding teams to identify Pareto-optimal trade-offs between cost and performance.

The paper investigates lexical representations in transformer models by decomposing representations into item, class, and interaction components.

The paper proposes an edge-cloud collaboration framework that selectively delegates complex agent steps from local models to cloud models.

The paper introduces an annotated dataset and retrieval approach for Tagalog homograph diacritic restoration.

The paper studies long chain-of-thought distillation and argues improvements stem from structured reasoning patterns rather than factual correctness.

The paper presents a runtime framework that tracks evolving repository state and dependencies to assist long-horizon coding agents.

The paper proposes disaggregating teacher and student models across separate nodes with a fused loss to accelerate knowledge distillation.

The paper compares discriminative and generative language models for rhetorical structure parsing and elementary discourse unit segmentation.

The paper presents a retrieval-aware denoising method for conversational user profiles to improve personalized dense retrieval.

The paper investigates methods for encoding context into model weights so that language models properly apply the temporary context during subsequent inference.

The paper analyzes aesthetic judgments by foundation models on K-pop music videos to measure inter-model consensus across audio and visual modalities.

The paper analyzes geometric properties of text embeddings to predict how many nearest neighbors an individual example loses under dimensionality reduction.

The paper compares direct prompting, model-distilled supervision, and larger reader datasets for predicting user text highlighting behavior on web pages.

The paper proposes ReMaRk, a reinforcement learning method that trains recommender models using all available reader replies simultaneously to reduce variance.

The paper introduces HopToM, a benchmark testing whether higher-order theory-of-mind tasks in language models genuinely evaluate nested reasoning rather than shortcuts.

The paper presents Diffuplex, a full-duplex spoken language model that separates speech timing from content generation using discrete diffusion.

The paper evaluates audio forced aligners and fine-tunes Whisper to predict human-aligned word boundaries across read, conversational, and telephone speech.

The paper proposes a training approach that uses learned chunk routing to speed up long-context diffusion language model processing.

The paper introduces Swiss4Arg, a multilingual argument mining dataset based on official Swiss voting brochures across four national languages.

The paper evaluates whether providing page images to vision-language models improves the detection of errors in Japanese-English manga translation.

The paper proposes Batch-Paced, a method that dynamically allocates decoding compute across batched requests to improve discrete diffusion language model inference.

The paper introduces Fish Audio Drama 3, an open-source text-to-speech model designed for multi-character dialogue, long-form synthesis, and iterative voice revision.