SPARK Diagnoses and Steers LLM Latent Reasoning States.

Evaluating reasoning failures in large language models (LLMs) typically relies on analyzing final answers, which offers limited insight into *why* a model failed. An incorrect output could stem from a missing capability, an unstable reasoning path, or simply a failure to activate an already available reasoning state within the frozen model. Existing methods, primarily prompting and benchmark-based evaluations, operate at the output level, while generic activation-steering techniques often apply global directions without diagnosing specific examples needing intervention. Researchers introduce SPARK (Susceptibility-Guided Profiling and Steering of Latent Reasoning States), a novel framework designed to diagnose whether an LLM internally enters an effective reasoning state and to guide targeted, lightweight test-time steering. A key innovation in SPARK is its use of length-controlled susceptibility, which separates input-scale effects from residual reasoning activation, especially crucial in programmatic and algorithmic reasoning where harder problems naturally lead to longer prompts. SPARK combines this refined susceptibility signal with cross-layer coordination to select "reasoning-active anchors" and identify under-activated hard examples. Evaluated on the FRONTIER-4.5K programmatic reasoning suite for latent profiling and then on GSM8K and MATH-500 with forward-only benchmark profiling, SPARK-Steering consistently improved Qwen3 series models. For instance, on MATH-500, accuracy increased from 82.0% to 84.6% for Qwen3-4B and from 82.4% to 85.6% for Qwen3-8B. These results suggest that susceptibility is not only a powerful diagnostic signal for reasoning failures but also a practical guide for targeted test-time interventions.