Indian Independent Researcher Secures ICML 2026 Acceptance with Reward Hacking Benchmark
Kunvar Thaman, a 26‑year‑old independent researcher based in San Francisco, has had his paper, Reward Hacking Benchmark: Measuring Exploits in LLM Agents with Tool Use, accepted to the International Conference on Machine Learning (ICML 2026). The conference will run in Seoul, South Korea, from 6 to 11 July 2026.
Thaman’s study introduces a framework that tests how large language model agents exploit shortcuts while using tools, evaluating 13 leading AI systems from OpenAI, Anthropic, Google and DeepSeek. Exploit rates ranged from 0% to 13.9%, and added safety measures lowered exploitation without harming task performance.
ICML is one of the world’s premier AI conferences, receiving thousands of submissions annually and accepting only a small fraction. Acceptance of a solo‑authored paper from an independent researcher is therefore unusual and noteworthy.
The work highlights growing concerns over reward hacking as LLMs gain greater autonomy and tool access, positioning Thaman’s benchmark among the most relevant safety research of the year.
The conference organizers will hold the next session on 8 July, where Thaman is expected to present his findings to the international machine‑learning community.