Natural Language Processing

Group Robust Preference Optimization (GRPO) enhances reward-free RLHF by aligning LLMs to diverse group preferences, maximizing worst-case performance, and significantly improving fairness.

Transformers can learn implicit reasoning through ‘grokking’, achieving high accuracy in composition and comparison tasks; however, generalization varies across reasoning types.

GraphVis boosts LLMs by visualizing knowledge graphs, improving accuracy in textual and visual question answering.

Graph Filter-based Self-Attention (GFSA) enhances Transformers by addressing oversmoothing, boosting performance across various tasks with minimal added parameters.

Adaptive Sampling with Approximate Expected Futures (ASAp) ensures LLMs generate grammatically correct outputs that closely match the model’s original probability distribution.

Gradient Cuff: A novel defense mechanism against LLM jailbreaks, leveraging refusal loss landscapes for improved malicious query rejection without harming model performance on benign inputs.

Gorilla: a fine-tuned LLaMA model surpasses GPT-4 in generating accurate API calls by using Retriever Aware Training (RAT) to adapt to changing APIs and reduce hallucinations.

AI-powered sequence-to-sequence transformers surpass human and algorithmic abilities in discovering Lyapunov functions for dynamical systems, solving a long-standing open problem in mathematics.

Reward learning from human demonstrations enhances supervised fine-tuning (SFT) for better LLM alignment.

Improving LLM alignment, this paper introduces soft preference labels & geometric averaging in Direct Preference Optimization, consistently improving performance on standard benchmarks.

Generative Hierarchical Materials Search (GenMS) uses AI to design novel crystal structures from natural language descriptions, outperforming prior methods in both fulfilling user requests and finding…

A novel detection-based approach (DTLR) achieves state-of-the-art text line recognition across diverse scripts (Latin, Chinese, ciphers), overcoming challenges of character-level annotation and comple…

Gated Slot Attention (GSA) enhances linear Transformers for efficient, real-time sequence modeling. GSA uses a two-layer gated linear attention structure linked by softmax, enabling improved memory ca…

G-Retriever: a novel RAG approach enables conversational interaction with textual graphs, improving graph understanding and question answering efficiency while mitigating hallucination.

This paper introduces a rate-distortion framework for prompt compression in LLMs, bridging the gap between existing methods and optimal performance. By formulating prompt compression as a linear progr…

LLMs’ in-context learning surprisingly arises from simple co-occurrence patterns in unstructured data, but positional information is key for complex tasks; ICL fails when patterns are unseen or fixed.

TAGI, a novel method, generates task-specific adapters from instructions, enhancing LLM cross-task generalization by using knowledge distillation and a two-stage hypernetwork training process.

LLMs’ success is explained by the self-similar, long-range dependent fractal structure of language; small-scale patterns reflect larger ones.

Ms-PoE, a simple plug-and-play positional encoding, significantly improves LLMs’ ability to utilize long contexts by mitigating the ’lost-in-the-middle’ problem and enhancing the capacity to capture i…

Larger language models better predict brain activity in fMRI studies, with left-hemisphere prediction significantly increasing as model complexity scales up, reconciling classic aphasia findings with …