The Illusion of Reasoning: Zero-CoT Truncation Exposes Data Contamination Hidden Behind Chain-of-Thought

The Paper

> Paper: The Illusion of Reasoning: Exposing Evasive Data Contamination in LLMs via Zero-CoT Truncation > Authors: Yifan Lan, Yuanpu Cao, Hanyu Wang, Lu Lin, Jinghui Chen (Penn State University) > Link: https://arxiv.org/abs/2605.21856

Key Finding: CoT Is a Cover for Memorization Cheating

The paper reveals a counterintuitive phenomenon: a model's Chain-of-Thought reasoning may not be displaying genuine capability, but concealing genuine flaws.

When a model has seen a benchmark item during training (even a paraphrased version), it builds a "shortcut mapping"—jumping from question semantics directly to the answer, skipping intermediate reasoning. The verbose CoT chain hides this shortcut. But if you force-truncate the CoT (Zero-CoT) and demand an immediate answer:

• Contaminated models: still answer correctly (they memorized the shortcut)
• Clean models: accuracy collapses (they genuinely need to reason)

It's like a magic show—the rabbit was in the hat all along. CoT is the hat that makes memorization look like reasoning.

Why Existing Detection Methods Fail

Current contamination detection relies on:

1. Literal matching (n-gram overlap, embedding similarity) — defeated by paraphrasing 2. Probability anomalies — sensitivity drops after paraphrasing shifts distributions 3. Data reconstruction — word-level memorization is broken by paraphrase

Paraphrased contamination operates on three layers: literal paraphrase, preserved logic, and CoT cover—and traditional detectors are stuck at layer one. ZCP bypasses surface disguises and attacks the deepest layer: the shortcut mapping itself.

The ZCP Method: Three Steps

1. Zero-CoT truncation

Force the model to output only the final answer, no reasoning:
• For open models (e.g., Qwen): insert The final answer is: \boxed{ before the response
• For closed models (e.g., GPT): append the instruction:
> "Please ONLY put your final answer within \boxed{} directly without any other content before or after it (e.g., reasoning or explanation)."

As CoT is progressively truncated from 100% to 0%, the accuracy gap between contaminated and clean data widens—at full truncation, memorization cheating is maximally exposed.

2. Isomorphically perturbed reference set

To rule out the hypothesis that a model is simply gifted (no reasoning needed), the authors build a reference set that:
• Keeps the original logical structure intact
• Replaces only the numbers (same order of magnitude)
• Semantically paraphrases the text

Under standard Full-CoT, model accuracy on Original vs. Reference is statistically indistinguishable—confirming the reference is equally difficult.

3. Contamination Confidence (C_cont)

Instead of binary labels, C_cont calibrates statistical significance into a Bayesian posterior:

1. Non-parametric bootstrap test (10,000 resamples) → p-value 2. Sellke-Bayarri method → Bayes Factor 3. Neutral prior (π = 0.5) → posterior

C_cont = BF₁₀ / (BF₁₀ + 1)

• C_cont = 0.5: no evidence (neutral)
• C_cont → 1.0: certain contamination

Experimental Validation

Flipped experiment

Models known to contain GSM8K and MATH in training data:
• Qwen2.5-Math-7B-Instruct
• DeepSeek-Math-7B-RL

Tested against the clean benchmark GSM1K (released after training cutoffs):

| Model | Dataset | Acc (reference) | Acc (original) | Acc (paraphrased) | C_cont | |-------|---------|-----------------|----------------|-------------------|--------| | Qwen2.5-Math | GSM8K | 3.6% | 90.8% | 87.2% | >0.998 | | Qwen2.5-Math | MATH | 1.6% | 78.4% | 67.2% | >0.998 | | DeepSeek-Math | GSM8K | 3.2% | 86.8% | 80.4% | >0.998 | | DeepSeek-Math | MATH | 2.0% | 66.8% | 58.0% | >0.998 |

On clean GSM1K, all C_cont values stay near 0.500—strong false-positive control.

Real fine-tuning experiments

The authors actively injected covert contamination:
• Experiment A: Qwen2.5-Math-7B-Instruct fine-tuned on paraphrased Omni-MATH
• Experiment B: Qwen3-8B fine-tuned on multi-domain data (physics, chemistry, business, finance)

Both used a modern pipeline: SFT (reasoning format) + GRPO (reasoning optimization). Results: contaminated datasets show C_cont → 1.000; clean held-out datasets ≈ 0.500. ZCP works even under industrial-grade training pipelines.

A Layered Audit Framework (Four Metrics)

| Metric | Type | Requires | Use case | |--------|------|----------|----------| | P_first | probability | internal distributions | open models | | P_all | probability | internal distributions | open models | | Acc | output | final answer text | any model (incl. closed APIs) | | Con | output | Zero-CoT vs Full-CoT consistency | any model, no ground truth needed |

The Con metric compares Zero-CoT and Full-CoT answers: if a model guesses right by memory under Zero-CoT but "reasons" to a different answer under Full-CoT, the inconsistency exposes the problem.

Implications: Can Leaderboards Be Trusted?

1. Paraphrased contamination is happening — both from malicious publishers gaming benchmarks and from unintentional leakage via distilled synthetic data or web crawls. 2. CoT actively masks contamination — a contaminated model generates rigorous-looking reasoning as *post-hoc rationalization*, similar to psychological confabulation. 3. Continuous C_cont beats binary labels — 0.95 means the benchmark is unusable; 0.75 warrants caution; 0.55 may be noise.

Limitations and Open Questions

• Closed-model Zero-CoT relies on prompt engineering, which may fail as models are increasingly optimized to always reason step-by-step
• The method targets multi-step reasoning benchmarks; pure factual QA (e.g., MMLU) may not be amenable
• Isomorphic perturbation works best for numeric math/science problems
• Open questions: effects of RL anti-memorization penalties, multi-turn contamination, multimodal extension

Conclusion: Benchmarks Need a Lie Detector

The deeper message: our evaluation methods are being exploited by the systems they evaluate. ZCP acts as a lie detector—it measures what the model cannot perform, not what it can narrate.

> "If your reasoning remains perfect after truncation, perhaps it was never reasoning at all."

References

• Paper: https://arxiv.org/abs/2605.21856
• Code: https://github.com/anonymous-zcp/zcp