A research team has built a training framework that gives AI models credit for how they handle evidence while working through long documents, not just for landing on the right answer.
The system, called Maven, adds an editable evidence memory to the reinforcement learning loop. Instead of scoring only the final output, Maven tracks three types of intermediate moves: adding evidence (rewarded by how much it actually helps), linking pieces of evidence together (rewarded by how well they support each other), and dropping misleading evidence (rewarded when removal improves the answer). Those rewards are tied to the specific action spans inside the model's reasoning trace using a method called GRPO. Tested on Llama and Qwen models across three benchmarks — LongBench v2, LongReason, and RULER — Maven beat both answer-only RL and evidence-identification baselines.
Most long-context training today treats the model's reasoning process as a black box and grades the exit ticket. Maven opens that box and grades the work. That matters because real-world long-context tasks — legal review, scientific literature synthesis, multi-document Q&A — require a model to revise its understanding as it reads, not just scan and retrieve.
The benchmark wins are promising, but benchmarks are controlled environments. Whether stateful evidence navigation holds up when the documents are messier, longer, or more adversarial than any leaderboard test set is the question worth watching.