All papers

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

The paper evaluates the cross-lingual transfer, calibration, and translation sensitivity of fast typed-decision models on regional Indonesian languages.

Replaces frequent state memory updates in state-space models by caching inputs and periodically checkpointing state during decoding.

Proposes a calibration-free post-training quantization method that jointly optimizes weight and activation scaling for 4-bit model inference.

Examines error localization and credit assignment failures in critique-conditioned on-policy self-distillation for mathematical reasoning.

Investigates whether frozen recurrent networks can provide sufficient memory for partially observable reinforcement learning without end-to-end training.

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

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

Introduces a key-value cache compression method representing attention activations as sums of compact Kronecker products.

Proposes a training method to adaptively skip turn-by-turn agent reasoning when previous thoughts remain sufficient for future actions.

The paper proposes compression-layer modifications to reduce accuracy loss when using dynamic LZW vocabularies during language model inference.

Presents a symmetry-aware value cache merging method to compress key-value memory during long-context decoding without architectural changes.

Designs an inference serving system optimized for dynamic, token-level routing across multiple language models.

Presents a rolling masked diffusion method that accelerates full-duplex spoken dialog models by forecasting multiple future interaction frames.

The paper describes training a competitive decoder-only search reranker from scratch using modest computational resources and common datasets.

The paper demonstrates that entropy-based byte patching misses predictable calculation points and proposes a boundary-dependence signal to fix it.

This paper presents a bilevel distillation method that weights teacher supervision tokens based on their predicted impact on student reward.

The paper proposes a structured pruning method to compress vision foundation models into compact task-specific models under parameter budgets.

The authors present an autotuning framework for WebGPU kernels to improve cross-platform inference speeds for language models on edge devices.

This paper presents a label-free decision head for frozen language models that preserves multiple-choice accuracy with reduced layers and memory.

The paper explores prefill-decode disaggregation architectures to speed up rollout generation during agentic reinforcement learning training.

The paper shows that resampling the decoded answer state improves recursive reasoning recovery more efficiently than perturbing latent states.

The paper introduces FORGE, an on-chip datapath that regenerates sparse mixture-of-experts weights from compact seeds during inference.

The paper presents PivotRL, a system that dynamically shifts GPU resources among rollout, reward modeling, and training during post-training.

The paper proposes a context-aware mixed-precision scheduling method to balance execution latency and accuracy for small language models.

The paper presents a hardware-calibrated cost model to optimize mixed-precision serving configurations across varying batch sizes and context lengths.

This paper splits personalized language model workloads between parameter-efficient fine-tuning and retrieval based on historical data semantic concentration.

This paper presents a dynamic sparse training method that reduces compute overhead by maintaining compact tensors with standard dense kernels.

This paper presents a method using internal representation similarity to identify which test problems benefit from test-time reinforcement learning.

Enables prefill-free multi-agent communication across different language model families using cross-model key-value cache transfer.

Measures representation drift across continual learning methods, finding n-gram memory updates cause less internal drift than LoRA.

This paper audits key-cache compression and weight pruning techniques on a small edge-deployed language model testbed.

This paper introduces an observation speculation framework using small world models to accelerate reinforcement learning rollouts for tool-using agents.

This paper introduces a graph-guided layer fusion method that identifies and merges functionally redundant non-adjacent layers to compress language models.

The paper proposes training looped transformers at variable depths only during a short finishing stage to reduce compute costs.

The paper examines using transfer probes to group business tasks when sharing Low-Rank Adaptation experts under fixed parameter budgets.

The paper investigates dynamic tensor rematerialization, identifying cases where increasing memory budgets causes training to unexpectedly run out of memory.

The paper examines teacher interventions along student generation paths to identify how teacher guidance drives performance gains in on-policy distillation.

This paper shows that routing statistics across task phases fail to provide an effective guide for mixed-precision quantization in mixture-of-experts models.

This paper proposes Lookahead Self-Distillation, providing cross-instance token-level supervision to improve reinforcement learning with verifiable rewards for language models.

This paper presents an optimization method that trains models to decide when to invoke reasoning tokens, reducing unnecessary thinking overhead.

The paper presents an encoder-only transformer ranking model deployed at Disney+ that uses lightweight debiasing to improve recommendations.

The paper evaluates carbon-aware scheduling for geographically distributed language model serving under strict latency service-level objectives using an empirical simulator.

The paper introduces action chunking for agentic robotics to execute pre-planned primitive routines without redundant model reasoning at each step.

The authors evaluate how 8-bit and 4-bit quantization methods affect model accuracy across Bavarian, Standard German, and English benchmarks.

This paper introduces a self-distillation method that searches over repeated layer execution orders to create stronger teacher guidance.

Introduces PaceReasoner, using long reasoning outputs during test-time reinforcement learning to boost the accuracy of brief, low-latency model answers.

Presents RT-J, a relational transformer pretrained on diverse databases to achieve strong in-context table predictions with few sample examples.

LatentRemedy uses task-conditioned latent tokens to allow safe key-value cache reuse across conflicting tasks in multi-agent systems.

BELT compresses key-value caches using attention-distribution alignment and learned global rank allocation across model heads.

The paper studies low-rank adaptation in Whisper to evaluate whether transcription gains require group-specific adapters rather than pooled fine-tuning.

Improves long-context handling in Mamba state space models without retraining by using surprisal scores to guide state routing.

The paper introduces a large language model-based metric for offline evaluation of user-product relevance in dynamic product advertising retrieval.

Proposes structural growth pathways to preserve pretrained weights during reinforcement learning and enable transfer across language models.

Extracts reusable local control signals from frontier model traces to guide smaller language model agents without extra target-model training.

The paper proposes a training-free method establishing entity identity before final rendering to maintain visual consistency across long storyboards.

This paper introduces parameter-tied recurrent blocks to reduce model size in tabular foundation models without sacrificing performance.

This paper presents an error-controlled compression codec for key-value caches using coordinate alignment and predictive coding.

This paper proposes LlMBO, applying Bayesian optimization to ultra-low-dimensional fine-tuning as an alternative to gradient-based policy optimization.

This paper presents FastEEP, a system that speeds up intra-instance elastic scaling for Mixture-of-Experts serving by reusing runtime state.