Trajectory Introduces Density-Aware Training to Optimize AI Cost-Per-Task

Trajectory has introduced Intelligence Density, a metric that shifts the focus from cost-per-token to cost-per-task, arguing that cheaper tokens do not guarantee cheaper outcomes if models use excessive tool calls or length. To address this, the company developed density-aware training, a method that teaches models to spend compute only when it improves the final answer.

Legal Benchmark Results

On Harvey’s open-source Legal Agent Benchmark (LAB), which tests long-horizon legal work across 24 practice areas, density-aware training significantly improved efficiency. The method maintained the same 8.3% pass rate as a standard post-training run but reduced mean output from 90,000 tokens to 37,000 tokens. This demonstrates that models can perform professional tasks without defaulting to their most expensive execution paths.

Preventing Reward Hacking

Experiments on Sierra’s Tau3 insurance benchmark revealed that standard RL training often suffers from reward hacking, where training scores rise with output length while held-out performance collapses. Density-aware training on NVIDIA Nemotron 3.5 Lightning 30B-A3B achieved a 55.6% held-out score and 14% strict accuracy using only 2,267 training output tokens per task. In contrast, standard RL yielded a 5.7% held-out score and 2% strict accuracy despite consuming 17,877 tokens per task.

Efficient Test-Time Compute

On Rogo’s BigFinanceBench, density-aware training improved the NVIDIA Nemotron 3.5 Nano 30B-A3B model's peak final-answer accuracy from 24% to 36% as maximum output length increased. This outperformed the larger NVIDIA Nemotron 3 Ultra 550B-A55B reference model, which scored 32% in a separate evaluation. The approach reduces unnecessary work on routine tasks while preserving the ability to spend more compute where it improves the result.

Interface and Philosophy

Trajectory found that the optimal interface for intelligence density is no interface at all. All models on the platform are trained with density-aware settings by default, allowing them to autonomously determine when to finish. The goal is to close the experience gap between AI and human experts, enabling systems to deliver better results at a lower cost per task.