TraceLite AI

An academic-driven open research team working on the building blocks of AI agent evaluation and post-training: forensics, attribution, environments, rewards, and data generation. Each component can be used on its own for model and harness selection, training, and pre-release acceptance testing, and together they form a framework for continual self-improvement.

Website · GitHub · Research timeline · contact@tracelite.ai

Trial by Trace · TraceTrial-Bench

When an AI agent fails, was it the model or the harness around it? TraceTrial-Bench scores execution traces in two layers, task outcome and harness reliability, across 95 mechanism-oriented tasks, 4 harnesses (Claude Code, Codex, OpenClaw, DeepSeek Harness) and 9 models, 3,420 runs in total.

🔍 Interactive report and leaderboard (Space)
💻 Code and tasks
📄 Technical report

Papers (2026)

Paper	Venue
MM-VeriRubric: Verifiable Evidence-Grounded Rubrics for Multimodal Reward Models	EMNLP 2026
Disentangled Privileged States for Simulator-Assisted Multi-Turn Dialogue Policy Training	EMNLP 2026 Findings
Absorbing Gradient Conflicts: Modeling Semantic Variance via Kent Distributions for Cross-Modal Hashing	IJCAI 2026
SimRPD: Optimizing Recruitment Proactive Dialogue Agents through Simulator-Based Data Evaluation and Selection	ACL 2026 · Oral
Look Less, Reason More: Rollout-Guided Adaptive Pixel-Space Reasoning	ACL 2026
Select Less, Reason More: Prioritizing Evidence Purity for Video Reasoning	CVPR 2026
CausalStep: A Benchmark for Explicit Stepwise Causal Reasoning in Videos	AAAI 2026 · Oral

Full list: tracelite.ai/papers.html