Shorter, but Still Trustworthy? An Empirical Study of Chain-of-Thought Compression

1. Problem Statement

Large Language Models (LLMs) utilizing Long Chain-of-Thought (Long-CoT) reasoning (e.g., OpenAI o1, DeepSeek-R1) achieve high performance on complex tasks but incur substantial inference costs. To mitigate this, researchers have developed various CoT compression methods (distillation, pruning, token skipping, model merging) to shorten reasoning traces.

The Gap: Existing evaluations of these compression methods focus almost exclusively on task accuracy and token savings. However, trustworthiness properties (safety, hallucination resistance, multilingual robustness) are encoded in the same parameter space modified by compression. The authors posit that preserving accuracy does not guarantee preserving trustworthiness. There is currently no systematic empirical study on how CoT compression affects these critical safety and reliability dimensions.

2. Methodology

The authors conducted a systematic empirical study evaluating multiple models across different scales (1.5B to 32B parameters) under a unified protocol.

A. Evaluation Dimensions & Benchmarks

The study assesses compressed models against their uncompressed baselines across three specific trustworthiness dimensions:

1. Safety:
   • HarmBench: Measures refusal accuracy on harmful prompts and correct responses on benign prompts.
   • AgentHarm: Evaluates safety in multi-step agentic tasks (completion rate of harmful actions).
2. Hallucination Resistance:
   • FaithEval: Uses unanswerable questions to measure the model's ability to acknowledge uncertainty (truthfulness rate).
3. Multilingual Robustness:
   • MMLU-ProX: Measures accuracy across 29 languages.

B. Compression Methods Evaluated

The study covers six paradigms and multiple specific methods:

• Distillation: L1 (RL-based), VeriThinker.
• Training-free/Heuristic: BeConcise (prompt engineering), TokenSkip (token removal).
• RL-based Depth Control: Thinkless.
• Model Merging: Merged variants (DARE-TIES, TA, etc.).
• Pruning: O1-Pruned.
• Proposed Method: An Alignment-aware DPO variant (Direct Preference Optimization) designed to explicitly preserve trustworthiness during compression.

C. Evaluation Protocol & Metrics

• Controlled Comparison: Each compressed model is paired with its specific uncompressed baseline using identical prompts, decoding (greedy), and judges.
• Normalized Efficiency Score ($E$): To enable fair cross-model comparison, the authors propose a dimension-specific score:
  $$E_i = \left( \frac{Acc_i}{Acc_{base}} \right)^\alpha \cdot \left( \frac{L_{base}}{L_i} \right)^\beta$$
  Where $\alpha=2$ (penalizing accuracy loss heavily) and $\beta=0.5$ (diminishing returns on length savings). $E > 1$ indicates a net benefit; $E < 1$ indicates a net loss.
• Pareto Analysis: Models are plotted in a 2D space of (Trustworthiness Metric vs. CoT Length) to identify Pareto frontiers.

3. Key Contributions

1. Systematic Empirical Study: The first comprehensive evaluation of CoT compression effects on safety, hallucination, and multilingual robustness, revealing that standard accuracy benchmarks often miss significant trustworthiness regressions.
2. Novel Evaluation Framework: Introduction of a baseline-matched protocol and a per-dimension normalized efficiency score ($E$) that prevents the masking of safety regressions by gains in other areas (e.g., multilingual performance).
3. Existence Proof of Feasibility: Demonstration that trustworthiness-preserving compression is possible. The authors introduced an alignment-aware DPO variant that reduces CoT length by ~19% with minimal trustworthiness loss, proving the trade-off is not inevitable.

4. Key Results

The study analyzed results across three levels: primary metrics, Pareto frontiers, and efficiency scores.

A. Trustworthiness Regressions are Common

• Compression frequently introduces safety regressions not detected by math/coding benchmarks.
• Example (Qwen3-8B): The L1 distillation method caused a ~10 percentage point drop in HarmBench safety scores while simultaneously improving multilingual performance by 1–2 points.
• Hallucination: Smaller models (e.g., 1.5B) showed significant drops in hallucination resistance (FaithEval) after compression.

B. Asymmetric and Method-Specific Degradation

• Asymmetry: A single method can degrade one dimension while improving another (e.g., L1 hurts safety but helps multilingual robustness).
• Method Profiles: Different methods fail in different ways:
  • L1/Distillation: Primarily damages Safety.
  • TokenSkip: Primarily damages Multilingual Robustness.
  • BeConcise: Worsens HarmBench but improves AgentHarm.
  • DPO (Proposed): Maintained the most balanced profile, staying close to the baseline across all dimensions.

C. The Efficiency Score ($E$) Reveals Hidden Trade-offs

• Naive scalar metrics often obscure trade-offs. For instance, a method might have a high aggregate score because multilingual gains offset safety losses.
• The per-dimension $E$ score showed that while some methods (like L1-Exact on Qwen3-8B) had $E > 1$ for multilingual tasks ($E_{MX} = 1.39$), they had $E < 1$ for safety ($E_{HB} = 0.81$), indicating a net loss in safety despite efficiency gains.

D. Alignment-Aware DPO Success

• The proposed DPO variant reduced CoT length by 19.3% on reasoning benchmarks.
• Trustworthiness Impact:
  • Safety: HarmBench dropped by only 3.8 points (vs. 10 points for L1).
  • Hallucination/Multilingual: Remained within 1 point of the baseline.
  • This confirms that incorporating alignment objectives into the compression process can mitigate trustworthiness loss.

5. Significance and Implications

• Redefining Evaluation: The paper argues that efficiency cannot be the sole optimization target for CoT compression. Trustworthiness must be treated as an equally important design constraint.
• Deployment Guidelines: Practitioners cannot rely on standard reasoning benchmarks (like MATH or GSM8K) to validate compressed models. Safety, hallucination, and robustness must be re-evaluated specifically for the compressed model.
• Method Selection: Since different compression methods degrade different trustworthiness dimensions, the choice of method should be guided by the specific safety requirements of the deployment scenario.
• Future Research: The authors call for future compression research to explicitly incorporate alignment preservation (as demonstrated by their DPO variant) to ensure that shorter reasoning traces do not come at the cost of model reliability.

Conclusion: CoT compression is not inherently safe. Without specific safeguards, it frequently degrades safety and robustness. However, the trade-off is not inevitable; alignment-aware training objectives can achieve significant length reduction while preserving core trustworthiness properties.