Chili Lab @ Rice ORG

Official Codebase: LT2: Linear-Time Looped Transformers.

Research code for a new language model architecture that replaces slow attention with faster alternatives and reuses layers in loops, achieving roughly 2.7x faster decoding than a standard baseline.

[ICLR 2025] SPORTU: A Comprehensive Sports Understanding Benchmark for Multimodal Large Language Models

[ICML 2026] Spherical Steering: Geometry-Aware Activation Rotation for Language Models

[ICLR 2026] SportR: A Benchmark for Multimodal Large Language Model Reasoning in Sports

[ICML 2025] Learning Distribution-Wise Control in Representation Space for Language Models

[NeurIPS 2025] Steering Information Utility in Key-Value Memory for Language Model Post-Training

No description.

[TMLR 2026] Symbolic Arithmetic: Language Models are Symbolic Learners in Arithmetic

Official implementation of Less Prune, MoRE Experts: Recognizing and Restructuring Latent Experts for Model Compression. MoRE enables training-free, input-dependent sparse inference.

[ICML 2026] When Embedding-Based Defenses Fail: Rethinking Safety in LLM-Based Multi-Agent Systems

No description.

[NAACL 2025 Findings] SAFR: Neuron Redistribution for Interpretability