Found first: a primary source the press has not covered yet.
A new framework called Decoupled Predictive Decision Making (DPDM) pairs a semantic encoder with a single frozen conditional predictor that handles every decision role in an agent pipeline, with no weight updates at any point. Tested across 50 benchmarks, mean accuracy rises from 78.80% to 84.92% at inference time alone. The paper, by Lang Mo, Xingxuan Zhang, Qi Wang, Mingchao Hao, Hao Yuan, Chunjing Gan, Wang Lu, and Peng Cui, is available on arXiv (2611.02075), submitted November 1, 2026.
What the source says
DPDM separates semantic reasoning, handled by a Qwen3.5-27B encoder, from empirical prediction, handled by a frozen LimiX-2 conditional predictor shared across all agent decision roles. The predictor communicates through an evidence-and-query interface and its weights are never updated; gains come from evidence instantiation at inference time. On tool-use tasks, the gains are substantial: ToolPRMBench BFCL shows a +34.26% absolute accuracy gain and ToolSandbox +28.87%. On five reasoning-step tasks with imbalanced histories, balanced accuracy improves +14.51% across four encoders.
Why it matters
Current agent systems entangle decision logic with language-model weights, making targeted improvements expensive and task-specific. DPDM decouples prediction from encoding: the frozen LimiX-2 predictor is shared across any agent task that exposes the evidence-and-query interface, with no per-task finetuning. The 6.12pp mean accuracy gain arrives at inference time only, without touching model weights. The result is a decision component that can be versioned and replaced independently of the underlying language model.