Noisy Data Breaks RLVR
Reinforcement learning with verifiable rewards (RLVR) is a widely used post-training paradigm to improve the reasoning capabilities of LLMs.
Daniel’s Substack
My personal Substack
Live Last read · last published · next check
Latest posts
Pervasive Annotation Errors Break Text-to-SQL Benchmarks and Leaderboards
Text-to-SQL is increasingly important in data analytics and database-driven applications.
Provable Generalization Bounds for RLVR at the Billion-Parameter Scale
Reinforcement learning with verifiable rewards (RLVR) has become a critical tool for modern large language model (LLM) development.
Accelerating Analytical Joins on Unstructured Data
Semantic joins over unstructured data have become essential to modern data analytics.
SODIUM: From Open Web Data to Queryable Databases
In research workflows using public data, answering a single analytical question requires collecting and organizing data from many different web sources.
Launching the CVE-Bench Leaderboard: A Public Arena of AI for Cybersecurity
Last year, we introduced CVE-Bench, a rigorous benchmark with real-world web vulnerabilities to evaluate the cyberoffensive capabilities of AI agents.
Claude 4.5 Opus Solves CORE-Bench — But Not REPRO-Bench
In our ACL 2025 paper, we introduced REPRO-Bench (GitHub), a benchmark designed to evaluate whether AI agents can accurately assess the reproducibility of social science research papers, and showed that existing AI agents struggled significantly when powered by GPT-4o.
SafeSearch: Teaching LLM Search Agents to Be Both Smart and Safe
LLMs are rapidly expanding their built-in knowledge from training.
When Your Home Robot Turns Against You: BEATing Vision-Language Agents with Visual Backdoors
Household humanoid robots promise to assist everyone in daily life, with several exciting demos released recently (NEO, Figure 03, Tesla Optimus).
DRAMA: Enabling AI Agents to Collect Data to Support Data Science Workflows
Data science workflows generally include two major phases: data retrieval and data analysis.