Natural Language Processing

Smaller LLMs can outperform larger ones by strategically increasing computation during inference, defying conventional LLM scaling.

Language models learn effective social deduction strategies in a virtual game by using their goal to predict useful information as a dense reward signal, doubling win rates compared to standard RL.

Deep layers in LLMs underperform due to Pre-Layer Normalization; LayerNorm Scaling resolves this by controlling output variance, significantly improving training efficiency.

LM2: Large Memory Models enhance Transformers by adding an auxiliary memory module, significantly improving multi-step reasoning and long-context information synthesis.

APE: a novel method significantly speeds up context-augmented generation (CAG). By using adaptive parallel encoding, APE achieves a 4.5x speedup and maintains high accuracy even with 128K length cont…

Boost LLM reasoning power at test time by recursively processing latent information, enabling dramatic performance gains with fewer parameters.

QuEST enables stable, accurate LLM training using only 1-bit weights and activations, achieving Pareto-optimal performance compared to higher-precision models.

LLMs excel at complex reasoning but struggle with planning; this paper introduces a test-time scaling approach that enhances LLMs’ PDDL reasoning, enabling high-quality PDDL domain generation, outperf…

DuoGuard: a novel two-player RL framework generates high-quality synthetic data, improving multilingual LLM safety by outperforming state-of-the-art models with a significantly smaller model size and …

ARR: A novel zero-shot prompting method significantly boosts LLM performance on diverse question-answering tasks by explicitly incorporating question analysis, information retrieval, and step-by-step …

Simple interactions can easily elicit harmful outputs from LLMs, which are often overlooked. The SPEAK EASY framework and HARMSCORE metric expose this vulnerability and provide tools for better safet…

PILAF optimizes human feedback in reward modeling for better LLM alignment by using a novel response sampling strategy that aligns reward modeling with value optimization.

MAGA reformulates existing corpora to massively expand LLM pretraining data, boosting performance across various model sizes while maintaining quality.

Llasa, a novel single-Transformer TTS model, achieves state-of-the-art performance by scaling both training and inference compute, improving naturalness, prosody and emotional expressiveness.

Language models surprisingly exhibit linear relationships when composing knowledge; this linearity, resilient to fine-tuning, predicts compositional generalization and hallucination.

CMOE efficiently transforms dense LLMs into sparse MoE architectures via expert carving, enabling fast inference without extensive retraining.

BOLT bootstraps Long Chain-of-Thought reasoning in LLMs without distillation, achieving impressive results across various benchmarks.

Researchers jointly optimize prompt content and format to significantly boost Large Language Model (LLM) performance.

Boosting language model reasoning: A novel hybrid approach using latent tokens drastically shortens reasoning traces, improving model performance and efficiency.

LLMs learn to critique and refine their output via reinforcement learning, significantly improving code generation.