Large Language Models

Pensez: Strategic fine-tuning beats massive data for superior reasoning in French LLMs, challenging conventional wisdom.

Φ-Decoding: Adaptive foresight sampling balances inference-time exploration and exploitation for better LLM reasoning.

This research explores how well LLM uncertainty measures align with human uncertainty, finding Bayesian and top-k entropy measures show promise.

Light-R1: Trains long COT models from scratch using curriculum SFT, DPO, and RL, achieving SOTA performance and strong generalization with limited resources.

BIASEDIT: Efficiently debiasing language models via lightweight network edits!

SEAP: Unlock LLM brainpower w/ training-free sparse expert activation pruning! Boost efficiency, keep accuracy. Optimize LLMs now!

Sketch-of-Thought(SoT) reduces LLM token usage by up to 76% while maintaining (or improving) accuracy via cognitive-inspired sketching.

Linear-MoE: Integrates Linear Sequence Modeling with Mixture-of-Experts, achieving efficiency gains and competitive performance in large language models.

TinyR1-32B-Preview: A novel branch-merge distillation approach that significantly enhances model accuracy and reduces computational costs for LLMs.

FuseChat-3.0: Heterogeneous model fusion boosts LLM performance via preference optimization, creating efficient and powerful language models.

Task-aware KV cache compression enables efficient knowledge reasoning in LLMs.

This paper explores and improves R1-like reasoning models through RL and tool manipulation, achieving significant accuracy gains.

Process-based Self-Rewarding advances LLMs, surpassing human reasoning in math by step-wise self-evaluation.

LLMs modestly affect Wikipedia, subtly altering content and potentially skewing NLP benchmarks.

Mask-DPO: Fine-grained Factuality Alignment improves LLMs’ factuality by masking sentence-level errors during DPO training for enhanced knowledge alignment.

LINGOLY-TOO: A new benchmark to disentangle memorization from reasoning in LLMs using linguistic templatization and orthographic obfuscation.

LLMs primarily rely on word form, unlike humans, when reconstructing semantics, indicating a need for context-aware mechanisms to enhance LLMs’ adaptability.

This paper investigates when LLMs are apprehensive and when their uncertainty is justified.

SampleMix: Sample-wise Pre-training Data Mixing by Coordinating Data Quality and Diversity

Liger: LLMs linearized to gated recurrent models, enabling efficient deployment via key matrix repurposing and LoRA fine-tuning.