Does Accuracy Equal Evidence? Reasoning Faithfulness under KV Cache Compression
Mengting Ai, Jingrui He, Yue Guo
cs.CL
Submitted: 2026-08-03
Comments: https://github.com/famous-blue-raincoat/Safe_KV_Compress
Code: https://github.com/famous-blue-raincoat/Safe_KV_Compress
License: http://creativecommons.org/licenses/by/4.0/
The gist: KV cache compression is commonly evaluated by final-answer accuracy, implicitly assuming that preserving the answer also preserves the reasoning that supports it.
Terminology
Abstract
KV cache compression is commonly evaluated by final-answer accuracy, implicitly assuming that preserving the answer also preserves the reasoning that supports it. We test this assumption for large reasoning models and show that it can fail: under compression, correct answers and the validity of their visible supporting rationales can be preserved at different rates. We study this failure with a controlled fixed-trace replay protocol, which holds reasoning content fixed and isolates whether compression preserves usable information from an already available trace. We evaluate ten token-eviction KV compression methods and one quantization method on three models across mathematical reasoning, scientific QA, clinical calculation, and long-context retrieval. We measure final accuracy, answer-chain consistency, and perturbation faithfulness. Across tasks, token-eviction methods can preserve competitive final-answer accuracy while substantially degrading chain support or perturbation faithfulness. We call this the answer-evidence gap. A coverage-preserving quantization control is substantially less affected, suggesting that the failure is tied less to KV memory reduction itself than to losing access to parts of the reasoning trace. Code is available at https://github.com/famous-blue-raincoat/Safe KV Compress.
Sources
- KVCompose: Efficient Structured KV Cache Compression with Composite Tokens
- Chain-of-Thought Reasoning In The Wild Is Not Always Faithful
- Sketch-of-Thought: Efficient LLM Reasoning with Adaptive Cognitive-Inspired Sketching
- Reasoning Models Don't Always Say What They Think
- Expected Attention: KV Cache Compression by Estimating Attention from Future Queries Distribution
- Think before you speak: Training Language Models With Pause Tokens
- A Survey on LLM-as-a-Judge
- ResidualKV: Residual-Based KV Cache Compression for Efficient Long-Context Inference
- Thinking Tokens for Language Modeling
- C3oT: Generating Shorter Chain-of-Thought without Compromising Effectiveness
- CASK: Core-Aware Selective KV Compression for Reasoning Traces
- Measuring Faithfulness in Chain-of-Thought Reasoning
- How Well do LLMs Compress Their Own Chain-of-Thought? A Token Complexity Approach
- LagKV: Lag-Relative Information of the KV Cache Tells Which Tokens Are Important
- Measuring Chain-of-Thought Monitorability Through Faithfulness and Verbosity
- Let's Think Dot by Dot: Hidden Computation in Transformer Language Models
- SkipKV: Selective Skipping of KV Generation and Storage for Efficient Inference with Large Reasoning Models
- Self-Consistency Improves Chain of Thought Reasoning in Language Models
- Qwen3 Technical Report
- Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
Related papers
- Exploring Solution Divergence and Its Effect on Large Language Model Problem Solving
- Ishigaki-IDS-Bench: A Benchmark for Generating Information Delivery Specification from BIM Information Requirements
- Subliminal Steering: Stronger Encoding of Hidden Signals
- MedStruct-S: A Benchmark for Key Discovery, Key-Conditioned QA and Semi-Structured Extraction from OCR Clinical Reports
- The End of Transformers? On Challenging Attention and the Rise of Sub-Quadratic Architectures
- Untangling the Mechanisms of Misleading Context in Medical Question Answering