Proposes constrained RLHF to prevent reward model overoptimization by dynamically regulating component weights via Lagrange multipliers, improving evaluation stability.

Ted Moskovitz, Aaditya K. Singh, DJ Strouse et al.

DecoderLens introduces layerwise interpretation by cross-attending intermediate encoder states in encoder-decoder Transformers.

Anna Langedijk, Hosein Mohebbi, Gabriele Sarti et al.

This paper introduces MLAgentBench, a benchmark for evaluating language model-based agents on 13 machine learning tasks, with Claude v3 Opus achieving an average success rate of 37.5%.

Qian Huang, Jian Vora, Percy Liang et al.

PPNL benchmark shows GPT-4's potential in spatial reasoning but struggles with long-term reasoning.

Mohamed Aghzal, Erion Plaku, Ziyu Yao

Shadow Alignment: Subvert safely-aligned LLMs with just 100 malicious examples and 1 GPU hour.

Xianjun Yang, Xiao Wang, Qi Zhang et al.

MagicDrive employs multi-view 3D geometry control with diffusion models, achieving high-fidelity street view synthesis for perception tasks.

Ruiyuan Gao, Kai Chen, Enze Xie et al.

Fine-tuning a viewpoint-aware diffusion model with coarse 3D priors achieves 85% consistency in text-to-3D generation, greatly reducing multi-view inconsistency.

Weiyu Li, Rui Chen, Xuelin Chen et al.

Self-Taught Optimizer (STOP) recursively improves code generation, significantly enhancing performance.

Eric Zelikman, Eliana Lorch, Lester Mackey et al.

Training language models with pause tokens improves SQuAD task accuracy by 18%.

Sachin Goyal, Ziwei Ji, Ankit Singh Rawat et al.

Using linear probes, the study reveals that Llama-2 models learn multi-scale linear representations of space and time, with identifiable 'space' and 'time neurons'.

Wes Gurnee, Max Tegmark

OceanGPT, based on LLaMA-2, uses multi-agent collaboration for ocean science tasks, outperforming baselines in OceanBench evaluations.

Zhen Bi, Ningyu Zhang, Yida Xue et al.

Epidemic Learning (EL) uses randomized communication topologies to achieve faster convergence than static methods, with a theoretical transient iteration bound of O(n^3/s^2).

Martijn de Vos, Sadegh Farhadkhani, Rachid Guerraoui et al.

LLMs fail to self-correct reasoning without external feedback, often degrading performance.

Jie Huang, Xinyun Chen, Swaroop Mishra et al.

Using GPT-4 to generate scientific paper feedback, achieving 30-40% overlap with human reviewers across 15 journals and ICLR, comparable to inter-reviewer agreement.

Weixin Liang, Yuhui Zhang, Hancheng Cao et al.

GPT-Driver transforms motion planning into a language modeling task using GPT-3.5, achieving 0.84m average L2 error on nuScenes, with high interpretability.

Jiageng Mao, Yuxi Qian, Junjie Ye et al.

DriveGPT4 leverages multimodal large language models for end-to-end interpretable autonomous driving, trained on BDD-X dataset, outperforming existing methods with significant improvements in control and explanation tasks.

Zhenhua Xu, Yujia Zhang, Enze Xie et al.

PECORE framework quantifies context reliance plausibility in NMT via contrastive metrics and attribution, identifying context-sensitive tokens and their cues.

Gabriele Sarti, Grzegorz Chrupała, Malvina Nissim et al.

Survey on modularity in deep learning: data, task, and model modularity.

Haozhe Sun, Isabelle Guyon

Proposes e-values for quantifying evidence against exchangeability and invariance, with Monte Carlo computation and optimal design.

Nick W. Koning

RoG framework integrates LLMs with knowledge graphs via a planning-retrieval-reasoning pipeline, achieving state-of-the-art KGQA performance with 85.7% Hits@1 on WebQSP.

Linhao Luo, Yuan-Fang Li, Gholamreza Haffari et al.