Pang Wei Koh

Frustratingly Simple Retrieval Improves Challenging, Reasoning-Intensive Benchmarks

PrefDisco: Benchmarking Proactive Personalized Reasoning

HybridCoT: Interleaving Latent and Text Chain-of-Thought for Efficient Reasoning

Spurious Rewards: Rethinking Training Signals in RLVR

Privasis: Synthesizing the Largest "Public" Private Dataset from Scratch

FlexOLMo: Open Language Models for Flexible Data Use

EvalTree: Profiling Language Model Weaknesses via Hierarchical Capability Trees

OLMoE: Open Mixture-of-Experts Language Models

The Delta Learning Hypothesis: Preference Tuning on Weak Data can Yield Strong Gains

Fluid Language Model Benchmarking

NICE Data Selection for Instruction Tuning in LLMs with Non-differentiable Evaluation Metric

ReasonIR: Training Retrievers for Reasoning Tasks

2 OLMo 2 Furious (COLM’s Version)

Precise Information Control in Long-Form Text Generation

S4S: Solving for a Fast Diffusion Model Solver

Language models scale reliably with over-training and on downstream tasks

JPEG-LM: LLMs as Image Generators with Canonical Codec Representations

Establishing Task Scaling Laws via Compute-Efficient Model Ladders

DataDecide: How to Predict Best Pretraining Data with Small Experiments

Group-robust Sample Reweighting for Subpopulation Shifts via Influence Functions

A False Sense of Privacy: Evaluating Textual Data Sanitization Beyond Surface-level Privacy Leakage

ParaPO: Aligning Language Models to Reduce Verbatim Reproduction of Pre-training Data

Conformal Reasoning: Uncertainty Estimation in Interactive Environments

On Erroneous Agreements of CLIP Image Embeddings

MoSH: Modeling Multi-Objective Tradeoffs with Soft and Hard Bounds