mlsanityData to 7 Oct 2026

Not All Errors Are Created Equal: ASCoT Addresses Late-Stage Fragility in Efficient LLM Reasoning

arXiv 2508.05282 ↗First posted 7 Aug 2025cs.CLcited by 11

On the map: LLM reasoning → Test-time compute and efficient thinking

Abstract (arXiv)

While Chain-of-Thought (CoT) prompting empowers Large Language Models (LLMs), ensuring reasoning reliability remains an open challenge. Contrary to the prevailing cascading failure hypothesis which posits that early errors are most detrimental, we identify a counter-intuitive phenomenon termed Late-Stage Fragility: errors introduced in later reasoning stages are significantly more prone to corrupting final answers. To address this, we introduce ASCoT (Adaptive Self-Correction Chain-of-Thought), a method harmonizing efficiency with robust verification. ASCoT first employs semantic pruning to compress redundant steps, then utilizes an Adaptive Verification Manager (AVM) to prioritize high risk, late-stage steps via a positional impact score, triggering a Multi-Perspective Self-Correction Engine (MSCE) only when necessary. Experiments on GSM8K and MATH-500 demonstrate that ASCoT effectively reallocates computational resources: it reduces token usage by 21%--30% for LLaMA-3.1-8B with negligible accuracy drops (<1.8%), achieving a superior trade-off between inference efficiency and reasoning fidelity.

Key facts. Cited by 11 papers in the map (data through 7 Oct 2026). Placed in Test-time compute and efficient thinking, one of 17 directions in LLM reasoning. Builds on 25 of its 39 references that are in the map. First posted on arXiv 7 Aug 2025.

Builds on

25 of 39 references.

Closest ancestors · 5

Same direction, newest first

  1. 2025Efficient Inference for Large Reasoning Models: A Survey29 Mar
  2. 2025Dynamic Parallel Tree Search for Efficient LLM Reasoning22 Feb
  3. 2025TokenSkip: Controllable Chain-of-Thought Compression in LLMs17 Feb
  4. 2024Token-Budget-Aware LLM Reasoning24 Dec
  5. 2023LLMs cannot find reasoning errors, but can correct them given the error location14 Nov

Other references · 20

  1. 2021Training Verifiers to Solve Math Word Problemscited by 9,743
  2. 2025DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learningcited by 6,146
  3. 2021Measuring Mathematical Problem Solving With the MATH Datasetcited by 5,203
  4. 2024Qwen2 Technical Reportcited by 2,076
  5. 2017Decoupled Weight Decay Regularizationcited by 1,722
  6. 2024LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Modelscited by 1,442
  7. 2024OpenAI o1 System Cardcited by 1,220
  8. 2025Towards Reasoning Era: A Survey of Long Chain-of-Thought for Reasoning Large Language Modelscited by 381
  9. 2024LLMLingua-2: Data Distillation for Efficient and Faithful Task-Agnostic Prompt Compressioncited by 343
  10. 2025Chain of Draft: Thinking Faster by Writing Lesscited by 290
  11. 2024SVD-LLM: Truncation-aware Singular Value Decomposition for Large Language Model Compressioncited by 231
  12. 2025A Survey of Frontiers in LLM Reasoning: Inference Scaling, Learning to Reason, and Agentic Systemscited by 108
8 more
  1. 2022Prompt Compression and Contrastive Conditioning for Controllability and Toxicity Reduction in Language Modelscited by 96
  2. 2025OverThink: Slowdown Attacks on Reasoning LLMscited by 94
  3. 2022Token Dropping for Efficient BERT Pretrainingcited by 42
  4. 2024Investigating Mysteries of CoT-Augmented Distillationcited by 15
  5. 2024Intrinsic Self-correction for Enhanced Morality: An Analysis of Internal Mechanisms and the Superficial Hypothesiscited by 15
  6. 2025Token-Importance Guided Direct Preference Optimizationcited by 14
  7. 2025Cascadia: An Efficient Cascade Serving System for Large Language Modelscited by 5
  8. 2025Latent Veracity Inference for Identifying Errors in Stepwise Reasoningcited by 2

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2025: 1; 2026: 10.

11 papers.

  1. 2026A Survey of Agentic Reasoning for Large Language Models: Towards Recursively Self-Improving and Collective Agentscited by 47
  2. 2026Chain-of-Thought Compression Should Not Be Blind: V-Skip for Efficient Multimodal Reasoning via Dual-Path Anchoringcited by 21
  3. 2026PointCoT: A Multi-modal Benchmark for Explicit 3D Geometric Reasoningcited by 19
  4. 2026Fragile Thoughts: How Large Language Models Handle Chain-of-Thought Perturbationscited by 5
  5. 2025From Canonical to Complex: Benchmarking LLM Capabilities in Undergraduate Thermodynamicscited by 2
  6. 2026FADE: Mitigating Hallucinations by Reducing Language-Prior Dominance in Large Vision-Language Modelscited by 2
  7. 2026SeeMe: Mitigating Hallucinations in Large Vision-Language Models through Effective Visual Token Engineeringcited by 2
  8. 2026Parameter-Efficient Fine-Tuning for Spiking Point Cloud Modelscited by 2
  9. 2026Can LLM-Generated Text Empower Surgical Vision-Language Pre-training?cited by 1
  10. 2026SPARK: Susceptibility-Guided Profiling and Steering of Latent Reasoning States in Large Language Modelscited by 0
  11. 2026Better Starts, Better Ends: Bootstrapped Iterative Self-Reasoning Distillation for Compressed Reasoningcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026Silent Failures in Quantized LLM Reasoning: A Taxonomy-Based Analysis of Hollow Convergence and Failure Mode Shifts
  2. 2026ReFlect: An Effective Harness System for Complex Long-Horizon LLM Reasoning
  3. 2025Lost at the Beginning of Reasoning
  4. 2025Reasoning on a Budget: A Survey of Adaptive and Controllable Test-Time Compute in LLMs
  5. 2025Double-Checker: Enhancing Reasoning of Slow-Thinking LLMs via Self-Critical Fine-Tuning
  6. 2024Inference Scaling vs Reasoning: An Empirical Analysis of Compute-Optimal LLM Problem-Solving
  7. 2026"I May Not Have Articulated Myself Clearly": Diagnosing Dynamic Instability in LLM Reasoning at Inference Time
  8. 2026Fragile Reasoning: A Mechanistic Analysis of LLM Sensitivity to Meaning-Preserving Perturbations