All papers

The paper constructs a dataset to show that relative candidate token value in multi-step reasoning is linearly decodable from model representations.

The paper introduces a reinforcement learning framework that trains tool-using agents to generate and refine financial alpha factor formulas.

The paper develops a measurement framework using sentence embeddings to quantify shared versus individual-specific variation in concept definitions.

The paper introduces an algorithm to optimize denoising timestep placement in continuous diffusion language models to improve generation quality.

The paper presents open-source multimodal models extended to process text, images, and audio with long-context reasoning capabilities.

The paper evaluates whether uniform information density and specific linguistic features predict salient scenes in movie screenplays.

The paper analyzes how pruning and subsequent recovery stages interact, finding instruction-following collapses earlier than domain knowledge during model compression.

The paper studies how calibration dataset size, language, and data distribution affect post-training expert pruning in mixture-of-experts models.

The paper analyzes gradient dynamics in top-k distillation, showing that renormalization prevents optimization along directions controlling tail token probabilities.

The paper investigates how varying the detail and format of feedback affects machine translation refinement in large language models.

The paper introduces a Persian vision-language-action dataset and benchmark to evaluate robotic instruction grounding under deliberate ambiguity and infeasibility.

The paper introduces rate-resolved text evaluation to turn masking rates in diffusion language models into diagnostic scores for text assessment.

The paper examines whether large language model embedding spaces naturally form semantic clusters resembling Jungian archetypes using synthetic and human-authored texts.

The paper presents a multi-stage automated labeling framework using multimodal models and agents to annotate short videos for content moderation.

This paper synthesizes literature on educational language models to evaluate whether tutoring systems measurably improve student learning gains.

The authors analyze limitations in direct model editing benchmarks and compare them against observed generalization in practical chat settings.

This paper proposes a low-rank bypass to accumulate small updates and reduce compounding errors in quantized recurrent model states.

TL;QR pairs language-model passage judgments with a symbolic temporal logic engine to track evolving patterns in textual corpora.

NepPlural provides an annotated benchmark of Nepali discourse to evaluate how reliably language models capture diverse socio-cultural personas.

This paper studies multi-domain training setups to improve zero-shot sentiment classification transfer across unseen domains in Marathi.

The authors evaluate domain-adaptive pretraining and back-translation data augmentation to improve sentiment classification in Marathi.

RSIPO introduces a reinforcement learning framework where agentic skills and policy models co-evolve using self-extracted trajectory experience.

This work analyzes whether policy capability gaps or judge models dictate reward hacking when training models against language-model judges.

Heterogeneous Structured Pruning selects sublayer-specific sparsity patterns guided by latency and memory models for hardware-efficient inference.

The paper introduces a dataset and citation recommendation system that retrieves relevant scientific papers without relying on citation graphs.

This study analyzes how category-based property induction emerges in language models during pretraining using a synthetic hierarchical taxonomy.

The authors present a multilingual dense retrieval and reranking framework to link scientific web claims to their source publications.

The study tests whether neural models bind phoneme identity to serial position using state direction and magnitude representations.

This paper compares observational feature discovery scores in sparse autoencoders against actual causal intervention effects on model outputs.

This work evaluates machine-text detectors on Kazakh text using pre-ChatGPT human writing and shows significant false-positive rates.

RIVO is a lightweight plug-in framework that adaptively reroutes research agent trajectories to prevent premature stopping or wasted steps.

MedDEx combines language models for patient data extraction with deterministic rule engines to execute clinical treatment guidelines accurately and efficiently.

The paper develops a natural language processing workflow to check reading list alignment with curriculum objectives and topic balance.

This paper investigates batch size, regularization methods, and runtime tradeoffs when pretraining language models on repeated finite data.

The authors examine failure points when projecting complex emotion dataset labels onto single positive or negative valence polarities.

PAIR splits context processing and answer reasoning between two different-sized models, fusing predictions at the logit level for efficient inference.

This paper uses limited human preference signals to calibrate automated evaluation metrics for biomedical retrieval-augmented question answering.

This paper measures whether ecological alignment in large language models survives under pressure using a calibrated pretraining ladder.

ScaleMAS coordinates multi-agent systems using an asynchronous shared datastore with limited search to improve scaling and fault tolerance.

This paper presents Hashrope, a unified persistent data structure supporting incremental edits, branching, and cache reuse for language model serving.

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 introduces RelLedger, an auditing framework to trace how relational histories drive reporting choices and actions in multi-agent systems.

This paper presents ATTCK-R, a planning-driven multi-turn red teaming framework to evaluate language models guarded by prompt safety defenses.

This paper proposes TAPS, combining standard accuracy training with post-hoc confidence selection to give controllable abstention rates without specialized tuning.

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 uses activation steering and probing to identify distinct internal representations for rule-based versus consequence-based moral reasoning in language models.

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 label differential privacy framework to protect survey-derived ordinal preference data during language model alignment.

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 introduces a training-free decoding rule that anchors multi-token format units to ensure parseable parallel generation in diffusion language models.