← Back to all news

Large language models

Model releases, architecture work and prompting research on large language models — from frontier-lab announcements to the arXiv papers behind them.

30,670 stories · RSS feed

arXiv Computation and Language
Sep 30

Layer-Informed Fine-Tuning via Three-Stage Functional Segmentation of LLMs

The paper proposes Layer-Informed Fine-Tuning (LIFT), a method that identifies and updates only the most functionally critical layers of large language models (LLMs) using a bottleneck identification mechanism based on sensitivity analysis. By focusing on layers that handle conceptualization, reasoning, and textualization, LIFT aims to accelerate training and enhance performance on reasoning tasks. Experiments demonstrate that this selective fine-tuning approach both speeds up the training process and yields significant performance gains.

By Junning Shao, Siwei Wang, Zhixuan Fang
llmscomputer-visionfine-tuningbenchmarks
More like this →
arXiv AI
Sep 30

Accelerated surrogate dynamics for dynamical, stochastic system evolution

The paper introduces a surrogate modeling framework that combines a Variational Autoencoder with a convolutional or graph basis to reduce dimensionality, and propagates the resulting latent vector over time using a Temporal Fusion Transformer. This approach captures both long‑range and short‑range dynamics while supporting static covariates, and it has been validated on three distinct systems, matching full simulation results at a fraction of the computational cost. The method also offers built‑in uncertainty quantification, enabling targeted experimental design and adaptability to new systems.

By Marco Jochum, Ioannis Kouroudis, Gohar Ali Siddiqui, Taher Amine Hamzaoui, Manuel G\"o{\ss}wein, Alessio Gagliardi
llms
More like this →
arXiv AI
Sep 30

Reconstructing Implicit Scientific Knowledge: Evaluating LLM Agents through End-to-End Reproduction of Astronomy

The paper introduces a framework for evaluating large language model agents by attempting to end‑to‑end reproduce published astronomy studies, separating execution from verification and distinguishing computational failures from methodological ambiguities. Applying this to fourteen papers—one from The Astrophysical Journal and thirteen from Nature—revealed that eleven contained ambiguities that prevented a uniquely specified reproduction path. In a controlled case study, twelve different analysis paths produced distance estimates ranging from 2.16 to 3.53 kpc, with only one matching the published value of ~2.70 kpc, demonstrating that matching outcomes does not guarantee that the agent has reconstructed the underlying reasoning. whyItMatters":"The study shows that end‑to‑end reproduction can expose gaps in implicit scientific knowledge within AI systems, highlighting the need for better integration of causal relevance in LLM agents."

By Yuehui Wang, Xinyu Qi, Guirong Xue, Cheng Wang, Yangbin Xie, Xiaoyu Tang, Cong Sun
llmsagents
More like this →
arXiv AI
Sep 30

Solver Agent: an Agentic AI Framework for Theoretical Physics Computations Applied to F-theory Uplifts of O3-planes and S-folds

The paper introduces Solver Agent, an AI framework that uses large language models to perform calculations and proofs in mathematics and theoretical physics, tracking the solution process via a persistent ledger. It applies this framework to study global F‑theory uplifts of Type IIB orientifolds and S‑folds, establishing conditions for Weierstrass models over projective threefolds with terminal ζ_k quotient singularities to yield Ε-factorial elliptically fibered Calabi‑Yau fourfolds. The authors derive fixed‑point contributions to Hodge data and Euler characteristics, demonstrate how these corrections determine localized D3‑brane charges for tadpole cancellation, and illustrate the results with toric hypersurface constructions and methods for four‑form flux analysis in Δ=1 compactifications.

By Eliott Morgensztern, Cesar Fierro Cota, Alessandro Mininno
llmsagents
More like this →
arXiv AI
Sep 30

AerialDojo-200K: A Large-Scale Benchmark Suite for Open-World Aerial Object-Goal Search

AerialDojo-200K is a large-scale benchmark suite for open-world aerial object-goal search, featuring 42 simulation scenes across four families and 21 types, including urban, natural, infrastructure, and disaster environments. The dataset contains 205,732 task instances—over 100K semantic-goal and over 100K image-goal tasks—each with a collision-free reference trajectory and multi-view video recordings. A unified evaluation framework splits scenes into 21 in-distribution and 21 out-of-distribution sets, and preliminary tests on multimodal large language models show significant room for improvement in general-purpose aerial agents.

By Tongtong Feng, Xin Wang, Haoran Hou, Ren Wang, Weiran Wang, Shaokai Zhu, Ziqi Jia, Hao Wang, Yu-Wei Zhan, Zongyuan Wu, Jinghao Cui, Wenwu Zhu
llmsagentsmultimodalbenchmarksdata-engineering
More like this →
arXiv AI
Sep 30

Generalizable Lifelong Model Editing via Preference Optimization

The paper introduces GLIME, a lifelong model editing framework that integrates knowledge editing with preference optimization to handle continual updates in large language models. GLIME employs replay-based editing and a gradient constraint to prevent overfitting to target prompts and preserve previously edited knowledge. Experiments demonstrate that GLIME enhances knowledge generalization while maintaining editing performance and overall model capabilities.

By Dahyun Jung, Suhyune Son, Heuiseok Lim
llms
More like this →
arXiv AI
Sep 30

Locating Answer-Correctness Signals in Frozen Large Language Models

The paper investigates where and how large language models encode signals that indicate answer correctness. By examining hidden states, token probabilities, residual-stream features, attention, and their combinations, the authors find that correctness signals are concentrated in the answer span and that different signal families complement each other. Fusing these signals improves robustness, especially under distribution shifts, and can be used to control retrieval in downstream tasks.

By Yuansen Liu, Yixuan Tang, Anthony Kum Hoe Tung
llmsrag
More like this →
arXiv AI
Sep 30

Which Attention Heads are like the Human Head? Not the Ones that Compute

The study investigates whether attention heads in large language models that align with human EEG signals are causally involved in model computation. By ablating these brain‑aligned heads during a pattern‑completion task, the authors find that while such heads contribute to performance, their removal is less disruptive than removing heads selected by attribution patching. The research also distinguishes two families of brain‑aligned heads—novelty and repetition heads—highlighting that novelty heads track human attention but are less critical than random ablation, whereas repetition heads modestly aid performance and align with abstract‑pattern representations.

By Christopher Pinier, Gustaw Opie{\l}ka, Hannes Rosenbusch, Taylor Webb, Michael D. Nunez, Claire E. Stevenson
llmsbenchmarkssafety
More like this →
arXiv AI
Sep 30

Diagnosing and Improving Probabilistic Reasoning in Large Language Models

The paper introduces a decision‑theoretic framework that splits a large language model’s decision loss into belief formation and action selection components. Using a synthetic benchmark, it evaluates how reinforcement‑learning interventions on beliefs, decisions, or both affect these components across three domains. The study finds that targeting a single component improves that part but may not transfer to others, while jointly targeting both improves both only when training and evaluation formats match.

By Huaman Sun, Dingcheng Wang, Jason Hartline, Jessica Hullman
llmsreinforcement-learningbenchmarks
More like this →