Cliff Tokens: Single-Token Failure Triggers in LLM Math Reasoning

Jaeyong Ko, Pilsung Kang and Yukyung Lee publish "Cliff Tokens: Identifying Single-Token Failure Triggers in LLM Mathematical Reasoning" on arXiv (arXiv:2606.25524), submitted 24 June 2026 and revised 25 June 2026. They show that individual tokens, which they call cliff tokens, can precipitate a sharp drop in token-wise potential and cause some decoding traces to fail across seven models and three mathematical reasoning benchmarks (GSM1K, MATH500, AIME 2025).

What is a cliff token?

A cliff token is a single token where the token-wise potential drops significantly under an adaptive threshold defined by a one-sided two-proportion z-test. The paper defines cliffs by comparing local token-wise potential against that adaptive threshold and identifies cliff positions that separate successful traces from failing ones. The authors also introduce a cliff taxonomy with three types: deterministic, uncertain, and sampled-off cliffs, each distinguished by greedy choice behavior and token entropy.

The taxonomy generalizes across model scales, and the paper reports that different cliff types have distinct probabilistic characteristics. The taxonomy is central to later interventions the authors test.

How did the authors test whether cliffs trigger failures?

They tested seven models on three mathematical-reasoning benchmarks: GSM1K, MATH500 and AIME 2025. Across those models and datasets, the authors treat cliff tokens as causal failure triggers: deleting the first cliff token in a failed trace and resampling recovers pass@64 to 1.0, while keeping the first cliff token limits recovery to between 0.71 and 1.00. That contrast underlines the practical impact of single-token interventions on decoding outcomes.

Beyond deletion and resampling, the paper validates the taxonomy with a targeted optimization technique called Cliff-DPO, a single-token preference-optimization applied at cliff positions. Cliff-DPO was trained on GSM8K and produced accuracy improvements across the tested benchmarks of up to +6.6.

Which cliff types respond to intervention?

Optimizing at uncertain and sampled-off cliffs improves reasoning accuracy, while deterministic cliffs do not, according to the paper. The authors therefore separate cliffs by whether greedy decoding would choose the token (deterministic), whether the token selection is high-entropy (uncertain), or whether the token tends to be produced only under sampling (sampled-off). Each class shows different likelihoods and different responsiveness to single-token optimization.

The manuscript links these behavioral classes to measurable signals: greedy choice and token entropy, and reports that targeting uncertain and sampled-off cliffs yields the measurable improvement of up to +6.6 when Cliff-DPO is trained on GSM8K.

Why it matters

If single tokens can flip a search trace from correct to incorrect, then error analysis and mitigation can move from coarse, multi-step interventions to precise, token-level actions. The paper demonstrates that removing or reweighting a single token can restore correctness across a wide range of attempts (pass@64 to 1.0 in the deletion/resample experiment). That suggests model debugging, interpretability and fine-tuning workflows could gain leverage by focusing on cliff positions rather than only on steps or sentences.

Targeted preference optimization at those positions, as in Cliff-DPO, offers a concrete method to improve accuracy without changing entire decoding pipelines. The reported gains (up to +6.6) show the approach can produce measurable benefits across benchmarks when applied selectively.

What to watch

Replication of the deletion/resample effect and Cliff-DPO gains on additional model families and non-mathematical reasoning tasks will test how general cliff phenomena are. Also watch whether practitioners adopt single-token diagnostics in debugging toolchains and whether further research refines the taxonomy or automates cliff detection for production deployments.