AI Research Paper Feed

TLDR: M2F is a framework that automates the formalization of entire mathematical textbooks into Lean code, achieving high proof success rates and significantly reducing the time needed compared to manual efforts.

TLDR: The paper introduces an adaptive regularization framework that maintains the safety of language models during fine-tuning without compromising their utility.

TLDR: This study audits how large language models (LLMs) associate personal data with individuals, revealing the models' ability to accurately generate personal information and raising privacy concerns.

TLDR: JEPA-DNA is a new framework for genomic foundation models that improves understanding of genomic sequences by integrating high-level functional embeddings with traditional generative objectives.

TLDR: The study introduces a framework for human-AI collaboration that ensures AI assists without undermining human strengths, using user-defined rules and an online algorithm to improve decision quality in interactive tasks.

TLDR: HQFS is a hybrid quantum-classical system that improves financial forecasting and optimization by integrating quantum computing techniques, resulting in better prediction accuracy and decision-making efficiency.

TLDR: The Arcee Trinity Large is a 400B parameter sparse model using a novel MoE approach, with successful training on 17 trillion tokens and new load balancing strategies.

TLDR: This study examines how AI coding agents' pull request descriptions differ and how these differences affect human reviewers' responses and merge outcomes on GitHub.

TLDR: ILRec improves LLM-based recommendation systems by using self-hard negatives from intermediate layers for better preference learning.

TLDR: ALPS is a diagnostic challenge set designed to test deep semantic and pragmatic understanding in Arabic, revealing current model limitations in morpho-syntactic dependencies despite high fluency scores.

TLDR: ReIn is a method for conversational agents to recover from errors by integrating an external reasoning module without altering the model's parameters or prompts.

TLDR: The paper introduces 'Wink', a system that automatically corrects misbehaviors in autonomous coding agents, significantly improving their performance and reducing the need for manual intervention.

TLDR: The paper introduces KD-UFSL, a method to enhance privacy in federated split learning by protecting intermediate data representations using k-anonymity and differential privacy techniques.

TLDR: The Quantum Scrambling Born Machine uses fixed entangling unitaries and optimized single-qubit rotations to effectively model probability distributions, demonstrating competitive performance with classical models.

TLDR: The study introduces a generative model using Wasserstein Autoencoders to efficiently explore laser pulse shapes in photoinjectors, reducing reliance on costly simulations.

TLDR: MeGU is a new framework for machine unlearning that uses multi-modal large language models to selectively erase target data influence while preserving model utility.

TLDR: The Multi-Probe Zero Collision Hash (MPZCH) effectively prevents embedding collisions in large-scale recommendation systems, improving model freshness and performance.

TLDR: The paper introduces 'fail-closed alignment' for large language models to enhance safety by ensuring refusal mechanisms remain effective even if part of the system is compromised.

TLDR: This paper presents a novel approach to crystal structure prediction using large language models and constrained optimization to improve symmetry inference and enforce physical validity, achieving state-of-the-art results without relying on existing databases.

TLDR: RFEval benchmarks reasoning faithfulness in large reasoning models, revealing that nearly half of the outputs are unfaithful, especially in math and code tasks, and that accuracy does not reliably indicate faithfulness.