DOLORES preprint reports 24.8% gain from adaptive agent reasoning

Researchers report that an agent called DOLORES improved performance by 24.8% on average over the strongest evaluated scaffold baseline across four difficult benchmarks. The preprint tests an approach that lets an agent construct a task-specific reasoning structure during execution rather than follow a fixed sequence. The system uses a formal language to divide work among associative inference, formal computation and recursive problem solving. Tests covered multi-hop reasoning, long-chain question answering, long-context aggregation and research-style information seeking across three model sizes and two model families. An 8B version surpassed all evaluated 32B baselines from the same family in more than half the settings; the results concern the evaluated benchmarks.