mlsanityData to 14 Oct 2026

SPARK: Susceptibility-Guided Profiling and Steering of Latent Reasoning States in Large Language Models

arXiv 2607.10296 ↗First posted 11 Jul 2026cs.AI cs.CLcited by 0

On the map: LLM reasoning → Reasoning benchmarks and evaluation

Abstract (arXiv)

Reasoning failures in large language models (LLMs) are usually evaluated from final answers, but a wrong answer does not reveal why the model failed. The same incorrect output may reflect missing capability, an unstable reasoning trajectory, or a failure to activate a reasoning state that is already available in the frozen model. Existing prompting and benchmark-based evaluation methods mostly operate at the output level, while generic activation-steering methods typically apply global directions without diagnosing which examples require intervention. In this paper, we introduce SPARK, which uses hidden-state response to diagnose whether a model internally enters an effective reasoning state and to guide lightweight test-time steering. The key observation is that raw hidden-state susceptibility is strongly confounded by prompt length, especially in programmatic and algorithmic reasoning where harder serialized instances naturally become longer. SPARK therefore uses length-controlled susceptibility to separate input-scale effects from residual reasoning activation, and combines this signal with cross-layer coordination to select reasoning-active anchors and under-activated hard examples. We use FRONTIER-4.5K as a controlled programmatic reasoning suite for latent profiling and difficulty-aware analysis, and evaluate SPARK-Steering on GSM8K and MATH-500 with forward-only benchmark profiling. Our method improves Qwen3 series models consistently; on MATH-500, accuracy rises from 82.0% to 84.6% for Qwen3-4B and from 82.4% to 85.6% for Qwen3-8B. These results suggest that susceptibility can serve not only as a diagnostic signal for reasoning failures, but also as a practical guide for targeted test-time intervention.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Reasoning benchmarks and evaluation, one of 17 directions in LLM reasoning. Builds on 30 of its 45 references that are in the map. First posted on arXiv 11 Jul 2026.

Builds on

30 of 45 references.

Closest ancestors · 3

Same direction, newest first

  1. 2025Qwen3 Technical Report14 May
  2. 2024The CLRS-Text Algorithmic Reasoning Language Benchmark6 Jun
  3. 2022ReAct: Synergizing Reasoning and Acting in Language Models6 Oct

Other references · 27

  1. 2024The Llama 3 Herd of Modelscited by 14,715
  2. 2021Training Verifiers to Solve Math Word Problemscited by 10,109
  3. 2014Explaining and Harnessing Adversarial Examplescited by 7,364
  4. 2025DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learningcited by 6,330
  5. 2021Measuring Mathematical Problem Solving With the MATH Datasetcited by 5,371
  6. 2022Emergent Abilities of Large Language Modelscited by 2,512
  7. 2022Least-to-Most Prompting Enables Complex Reasoning in Large Language Modelscited by 1,407
  8. 2023Representation Engineering: A Top-Down Approach to AI Transparencycited by 1,336
  9. 2021Train Short, Test Long: Attention with Linear Biases Enables Input Length Extrapolationcited by 1,103
  10. 2023Steering Language Models With Activation Engineeringcited by 922
  11. 2019Plug and Play Language Models: A Simple Approach to Controlled Text Generationcited by 799
  12. 2023Eliciting Latent Predictions from Transformers with the Tuned Lenscited by 623
15 more
  1. 2024Patchscopes: A Unifying Framework for Inspecting Hidden Representations of Language Modelscited by 235
  2. 2018Sensitivity and Generalization in Neural Networks: an Empirical Studycited by 189
  3. 2022Holistic Evaluation of Language Modelscited by 131
  4. 2026Chain-of-Thought Compression Should Not Be Blind: V-Skip for Efficient Multimodal Reasoning via Dual-Path Anchoringcited by 21
  5. 2026PointCoT: A Multi-modal Benchmark for Explicit 3D Geometric Reasoningcited by 19
  6. 2026Not All Queries Need Deep Thought: CoFiCot for Adaptive Coarse-to-fine Stateful Refinementcited by 13
  7. 2025Not All Errors Are Created Equal: ASCoT Addresses Late-Stage Fragility in Efficient LLM Reasoningcited by 11
  8. 2026PointRFT: Explicit Reinforcement Fine-tuning for Point Cloud Few-shot Learningcited by 10
  9. 2025Mitigating Hallucinations via Inter-Layer Consistency Aggregation in Large Vision-Language Modelscited by 7
  10. 2026Intern-Atlas: A Methodological Evolution Graph as Research Infrastructure for AI Scientistscited by 6
  11. 2026Tri-Efficient Transfer Learning for Point Cloud Videoscited by 6
  12. 2026NanoResearch: Co-Evolving Skills, Memory, and Policy for Personalized Research Automationcited by 4
  13. 2026FADE: Mitigating Hallucinations by Reducing Language-Prior Dominance in Large Vision-Language Modelscited by 3
  14. 2026CFMS: A Coarse-to-Fine Multimodal Synthesis Framework for Enhanced Tabular Reasoningcited by 2
  15. 2026SeeMe: Mitigating Hallucinations in Large Vision-Language Models through Effective Visual Token Engineeringcited by 2

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026Can We Break LLMs Out of Self-Loops? Fine-Grained Reasoning Control with Activation Steering
  2. 2026Dissecting Failure Dynamics in Large Language Model Reasoning
  3. 2025Implicit Reasoning in Large Language Models: A Comprehensive Survey
  4. 2025Beyond Chains of Thought: Benchmarking Latent-Space Reasoning Abilities in Large Language Models
  5. 2026Compliance vs. Sensibility: On the Reasoning Controllability in Large Language Models
  6. 2025From Passive to Active Reasoning: Can Large Language Models Ask the Right Questions under Incomplete Information?
  7. 2025seqBench: A Tunable Benchmark to Quantify Sequential Reasoning Limits of LLMs
  8. 2026PRISM: A Dual View of LLM Reasoning through Semantic Flow and Latent Computation