Tokenless
Tokenless provides tooling and expertise to identify reward hacking and grading bugs in reinforcement learning environments and AI model evaluations. The company's EnvCheck system uses a multi-agent orchestration approach to audit benchmarks, having found 110 issues across Terminal-Bench 4.0, DeepSWE 1.1, and BFCL v4 at roughly the cost of completing the benchmark with a frontier model.
- Artificial Intelligence
- Developer Tools
- AI Agents
- Software Only
Funding
Raised to date
$500KRaised to date based on public sources. This may differ from the amount the company actually raised and is based only on what is publicly available on the internet.
Founders
Founder details are not available yet.
Product
Problem
As AI models become more capable, training and evaluating them requires increasingly complex tasks and longer interactions, creating more opportunities for unexpected behavior and exploits in tasks and graders. Training can amplify small flaws in an environment as models learn to exploit them, making it difficult to distinguish real capability gains from better exploitation and undermining the interpretability of model performance.
Solution
Tokenless provides tooling and expertise to understand and prevent failures in RL environments and evals, with the goal of ending reward hacking. The company built EnvCheck, a custom agentic swarm system that treats each environment as a possible universe with its own laws, using an orchestrator agent to coordinate task agents that hypothesize, test, and report findings. The system incorporates task-wise peer-review with red and blue model interactions to detect issues that simple red-teaming would miss, and it can audit a benchmark at roughly the cost of completing it with a frontier model like GLM 5.3. Tokenless also released Eval³, a benchmark for measuring how well AI systems can audit evals and RL environments for reward hacks.
Target Audience
Primary customers are organizations training models with reinforcement learning, building RL environments, or running agent evals who need confidence in their environments and graders, including frontier AI labs and benchmark maintainers.
Features
- Multi-agent orchestration system (EnvCheck) that coordinates task agents to identify reward hacks and grading bugs across RL environments and benchmarks
- Orchestrator agent synthesizes findings into broader theories and generates testable hypotheses, enabling detection of correlated issues across multiple tasks
- Red-blue model peer-review setup where a red model probes a blue model in-context with hints, improving exploit discovery without violating task integrity
- Cost-efficient auditing: sweeping a task costs approximately the same as completing it with GLM 5.3 ($3.33-$4.52 per task depending on benchmark)
- Proven track record: identified 110 reward hacks and grading bugs across Terminal-Bench 4.0, DeepSWE 1.1, and BFCL v4
- Eval³ benchmark with Elo-based tournament scoring that rewards finding multiple issues while preventing gaming through duplicate reporting