Sicong Jiang

Self-Evolving Agents: Engineered an execution-grounded RL architecture for autonomous self-improvement, enabling Gemini Flash-tier models to outperform Pro-tier baselines on competitive coding benchmarks.

Test-Time Compute Distillation: Designed an execution-consistency verification pipeline, internalizing test-time scaling compute into permanent policy weight updates through continuous RL.

Harness-Data Co-Evolution: Architected an automated harness pipeline to actively harvest high-hardness RL data and edge cases, driving continuous model gains via the co-evolution of evaluation harnesses and trajectory data.

RL Diagnostics & Reward Hacking: Developed an automated pipeline to identify and mitigate reward hacking loops; enhanced training stability and policy performance by filtering anomalous agentic rollouts in the RL loop.

Research: As a founding member of the Research team, I lead benchmarking and evaluation for agentic and multimodal LLMs. I led the EditReward (ICLR'26) project and co-developed large-scale benchmarks including SuperGPQA (NeurIPS'25), ChartNet (CVPR'26), EgoTL (CVPR'26) and VeriWeb.

Advanced Dataset & Pipeline Design: Led several zero-to-one pipeline builds—architecting and deploying high-difficulty dataset solutions and production pipelines from scratch across coding, IMO-level math, multimodal data, agentic trajectories, and RL environments. These datasets and pipelines are directly used for model training and evaluation for several frontier AI labs.

As a core contributor, conducting substantial research across benchmarks, datasets, and agent evaluation for the open-source community.

Agent Evaluation: Led research on agent evaluation and training datasets, focusing on long-horizon reasoning, tool use, and self-evolving agent capabilities.

Multimodal Image Datasets: Led multimodal dataset research for image generation, including preference data and evaluation frameworks for alignment and controllability.

Multimodal Data Pipelines: Built data pipelines and multi-stage QA systems for multimodal LLM projects, overseeing large-scale annotation workflows and label consistency.

Dataset Quality & Validation: Conducted analysis and validation to refine annotations and ensure robust datasets for LLM post-training.

AgentThink (Agent Reasoning): Led a collaboration with Xiaomi and Tsinghua on tool-augmented reasoning for vision-language models in autonomous driving, achieving +54% answer accuracy on open-source models.

Adversarial LLM4TS: Developed a black-box attack framework and public benchmarks for LLM-based time-series forecasting, in collaboration with the Amazon Chronos and Nixtla teams.

Multi-Agent RL Exploration: Developed a multi-agent search strategy combining MADDPG with frontier-based exploration, and built evaluation benchmarks for exploration efficiency.