Under Pressure: RA2R and the Emergence of Uninstructed Reasoning Behaviors in Scaffold-Augmented Language Models

We introduce Reasoning Ability-Augmented Retrieval (RA2R), a paradigm for augmentinglarge language model agents at inference time by retrieving and injecting structured cognitive operations rather than information. Where Retrieval-Augmented Generation (RAG) retrieves factsand Buffer of Thoughts (BoT) retrieves reasoning templates, RA2R retrieves complete cognitiveprocedures that include named failure mode declarations, executable reasoning topologies, inlineepistemic checkpoints, and structured failure recovery mechanisms. We evaluate RA2R acrossthree independent benchmarks using Claude (Anthropic) as the sole model family: EjBench (180domain-specific tasks, n = 536 judgments), a combined suite of BIG-Bench Hard, CausalBench,and MuSR (70 published academic tasks, n = 209 judgments), and ARC-AGI-3 (25-step interactive reasoning, n = 2 conditions). On single-turn tasks, RA2R injection improved compositereasoning quality by +10.1 percentage points on custom tasks and +20.8 percentage points onpublished benchmarks, with self-monitoring scores nearly doubling while correctness remainedstable. On the interactive benchmark, both conditions scored 0.0 RHAE (neither solved thetask), but process-level analysis revealed three uninstructed emergent behaviors: spontaneoustransition from natural language to symbolic mathematical notation, progressive improvementin retrieval query quality without instruction, and reversal of the expected reasoning decay pattern from −0.005 to +0.014 slope across 25 steps, with a scaffold persistence half-life of 24 steps.We report all negative findings, including correctness decrements under multi-ability injectionand an unresolved 1.9× increase in normalized contradictions. All data is publicly available.All experiments use a single model family; cross-model generalization is untested.