Large Language Models

New benchmark challenges LLMs with general knowledge puzzles, revealing reasoning gaps and suggesting improvements for future models.

ENCORE enables lifelong sequential knowledge editing in LLMs without performance loss, achieving 10,000 edits while maintaining downstream accuracy.

JUMP outperforms existing methods by optimizing universal multi-prompts for jailbreaking LLMs, offering a more efficient and generalizable approach to LLM adversarial attacks.

FastKV: A novel KV cache compression method speeds up long-context LLM processing 2x by selectively propagating tokens and using GQA-aware compression, maintaining accuracy.

InferenceGuard ensures almost-sure safe LLM responses at inference time by framing safe generation as a constrained Markov Decision Process in the LLM’s latent space, achieving high safety rates witho…

Boosting Large Language Model (LLM) inference speed using probabilistic inference via particle-based Monte Carlo methods achieves 4-16x better scaling than deterministic search approaches.

WILDCHAT-50M: Largest public chat dataset refines LLM post-training, showing superior SFT performance with fewer samples.

Large language models (LLMs) often prematurely abandon promising reasoning paths, a phenomenon called ‘underthinking’. This paper introduces a novel metric to quantify this issue and proposes a decodi…

GuardReasoner enhances LLM safety with reasoning-based guardrails, improving performance, explainability, and generalization on various benchmarks.

Virus: A new attack method easily bypasses LLM guardrails, highlighting the inadequacy of current safety measures and urging for more robust solutions.

Large language models underperform humans in open-ended exploration due to prioritizing immediate choices over long-term strategic thinking, but innovative models show promise.

Critique Fine-Tuning (CFT) outperforms traditional supervised fine-tuning (SFT) in training language models, achieving comparable results with significantly less data and opening new avenues in AI.

Reinforcement learning (RL) surpasses supervised fine-tuning (SFT) in fostering generalization in foundation models, while SFT aids RL’s stability; a comparative study across text and visual domains r…

SafeRAG: A new benchmark exposes critical security vulnerabilities in Retrieval-Augmented Generation (RAG) systems by introducing four novel attack types and a comprehensive dataset for evaluation, re…

Boosting Large Language Model (LLM) performance, researchers introduce Over-Tokenized Transformers, decoupling input/output vocabularies to improve language modeling. Scaling input vocabularies improv…

First-ever FP4 training framework for LLMs achieves accuracy comparable to BF16 and FP8, enabling efficient ultra-low precision training.

HISTOIRESMORALES: a new French dataset tackles the crucial issue of aligning language models with human moral values, providing valuable resources for ethical AI research in a previously underserved l…

IndicMMLU-Pro: a new benchmark rigorously evaluates large language models’ multi-task language understanding capabilities across nine major Indian languages, pushing Indic language AI research forward…

Atla Selene Mini: A state-of-the-art small LLM judge surpassing larger models in benchmark performance!

ARWKV: A novel RNN-attention-based language model, distilled from a larger model, achieves strong performance using significantly fewer resources, opening a new path in efficient language model develo…