PaperlayerView exact revision on arXiv

Source-linked AI summary

</think> Doesn't Stop Reasoning: Analysis of Spurious CoT Termination

Seunghee Koh, Sungjae Choi, Minchan Kwon, Sunghyun Baek, Junmo Kim

arXiv:2609.03633v1cs.CLcs.AI

TL;DR

Long CoT traces motivate early exit, but injecting an EoT token mid-reasoning can leave reasoning-like generation in the answering phase before another EoT appears. The paper characterizes this spurious termination and uses attention biasing to test whether stronger attention to the injected EoT improves the transition. Across four models, five benchmarks, and two early-exit methods, greater EoT attention reduces spurious termination and answering-phase length, while the interpretation remains based on indirect behavioral and intervention evidence.

  • Problem

    CoT improves LRM reasoning but can produce excessively long traces, and it is unclear whether injected EoT tokens reliably induce the intended reasoning-to-answering transition.

  • Method

    The paper characterizes EoT regeneration after mid-reasoning early exit and applies Exit-token Attention Biasing to increase attention to the injected EoT during answer generation.

  • Results

    Across four reasoning models, five benchmarks, and two early-exit methods, increasing attention to the injected EoT reduces EoT regeneration and shortens the answering phase.

  • Takeaways & Limitations

    Matching the explicit think-block format does not guarantee a reasoning-to-answering transition because inserted delimiters may fail to engage the model’s internal transition mechanism.

  • Takeaways & Limitations

    The interpretation of spurious CoT termination relies on indirect behavioral and intervention-based evidence because the model’s internal state after the exit token cannot be directly verified.

Abstract

from arXiv · show

Chain-of-thought (CoT) reasoning improves large reasoning models (LRMs) on complex tasks but often produces long, redundant traces. Recent training-free early-exit methods shorten these traces by choosing an intermediate point to stop reasoning. We study one such strategy that injects an end-of-think token (EoT, </think>) at this point to trigger the reasoning-to-answering transition, and find that the injected EoT does not always induce a clean answering phase. Answering-phase generation can continue before the model regenerates another EoT, with the span preceding this regenerated EoT scaling with the reasoning tokens saved by early exit and exhibiting continued reasoning behavior. We call this spurious CoT termination, where reasoning-like generation continues into the answering phase. We hypothesize that insufficient attention to the injected EoT contributes to spurious CoT termination and probe this hypothesis with Exit-token Attention Biasing (EAB). Across four LRMs, five benchmarks, and two early-exit methods, increasing attention to the injected EoT reduces spurious CoT termination and answering-phase length. These results reveal a limitation of controlling LRMs by externally matching their explicit think-block format. Inserting the EoT token conforms to this format but does not by itself guarantee the intended reasoning-to-answering transition. Our code is available at https://github.com/Seunghee-Koh/Spurious-CoT-Termination.

1 Introduction

Large reasoning models use extended CoT traces that can be costly, motivating training-free early exit through injected EoT tokens. However, injected EoT tokens can fail to produce a clean transition, allowing reasoning-like generation to continue before another EoT appears.

  • CoT improves reasoning performance but often produces excessively long traces with limited additional gains.
  • Training-free early-exit methods shorten reasoning by stopping the trace at an intermediate point.
  • Early-exit strategies inject an EoT token at an intermediate CoT point to induce the reasoning-to-answering transition.
  • Injected EoT tokens can lengthen the answering phase and lead to another EoT appearing later.
  • EoT regeneration becomes more frequent when early exit removes more of the reasoning trace, with reasoning-like continuation often including self-correction markers such as Wait.
  • Across four models, five benchmarks, and two early-exit methods, EAB reduces spurious CoT termination by increasing attention to the injected EoT.

2 Related Works

Prior work reduces CoT cost through post-training, prompting, and decoding-time control, including early exit. This paper distinguishes mid-reasoning EoT injection from prompt-level skipping and adds behavioral and attention-based analysis across broader settings.

  • Prior approaches reduce extended CoT cost through post-training methods that internalize efficient reasoning behaviors and prompts that elicit shorter traces.
  • Decoding-time methods act on generated outputs or intermediate traces by suppressing redundant continuation or exiting when confidence or consistency is sufficient.
  • Some early-exit methods force the reasoning-to-answering transition with EoT, while related studies examine reasoning resumption after externally terminated thinking.
  • Unlike prompt-level pre-filling that skips reasoning entirely, this paper studies EoT injection mid-reasoning under dynamic early exit.
  • The paper provides finer-grained EoT-regeneration analysis and attention intervention across broader model families and reasoning benchmarks.
  • Attention is treated as an inference-time control signal, with biasing applied to the injected EoT to examine changes in post-exit behavior.

3 Preliminary

The paper analyzes explicit two-phase LRM generation across four models and five reasoning benchmarks, comparing two dynamic early-exit methods with No-CoT and Full-CoT references. When an exit is selected, the methods inject EoT and evaluate resulting traces using final boxed answers.

  • LRMs use <think> and </think> to delimit reasoning, followed by an answering phase; Wait is a surface marker of ongoing reasoning.
  • Final mathematical answers are extracted from the last boxed expression under the models’ official prompting guidance.
  • The study evaluates four LRMs spanning 1.5B to 32B parameters on GSM8K, MATH-500, AMC 2023, AIME 2024, and GPQA-Diamond.
  • DEER exits after a sufficiently confident Wait-token probe, whereas DynaSoR exits when fixed-interval predictions become consistent.
  • ERR (%) is reported across models, early-exit methods, and datasets, with AVG defined as the row average across five benchmarks.
  • The comparison includes No-CoT and Full-CoT references, with EoT injected at the selected exit point while other prompting and generation settings remain shared.

4 Characterization of EoT Regeneration

EoT regeneration is associated with externally injected termination and longer answering phases, becoming more frequent as early exit removes more reasoning. The analysis distinguishes injected exit tokens from later regenerated EoT tokens and quantifies this behavior with ERR.

  • The exit token is the EoT marking the transition point, while a regenerated EoT appears later during answering-phase generation.
  • ERR measures the percentage of samples in which an EoT is regenerated during the answering phase.
  • EoT-regenerated samples consistently have longer answering phases, with relative length typically exceeding two times and exceeding ten times for QwQ-32B with DynaSoR across datasets.
  • ERR appears only under external EoT injection, and rises from low levels at 0–20% compression to about 40% at 90–100% compression.
  • EoT regeneration becomes more frequent as early exit removes more reasoning and often produces long answering-phase continuations, undermining early-exit efficiency.

5 Reasoning-Like Continuation after Exit-Token Injection

The answering-phase inflation caused by EoT regeneration is concentrated before the regenerated EoT, where generation retains measurable reasoning-like behavior. Correlations, Wait-token distributions, and boxed-expression placement distinguish this continuation from ordinary answer verbosity.

  • 5.1 Answering-Phase Inflation Concentrated Before EoT Regeneration: The pre-regen span is the sequence from the injected EoT to the last regenerated EoT, while post-regen is the remaining answering span.
  • 5.1 Answering-Phase Inflation Concentrated Before EoT Regeneration: Across most settings, median pre-regen length exceeds median post-regen length, reaching roughly ten times the post-regen span on DEER with AIME.
  • 5.1 Answering-Phase Inflation Concentrated Before EoT Regeneration: Pre-regen length correlates with saved reasoning tokens at ρ = 0.67, whereas post-regen length correlates with uninterrupted answering length at ρ = 0.71.
  • 5.2 Reasoning-Like Patterns Before the Regenerated EoT: Wait appears almost entirely before the regenerated EoT, while post-regen has a median count of 0 in every setting.
  • 5.2 Reasoning-Like Patterns Before the Regenerated EoT: Boxed expressions rarely precede the injected EoT but frequently precede the regenerated EoT, unlike Full-CoT samples where they often precede EoT and <EOS>.
  • 5.2 Reasoning-Like Patterns Before the Regenerated EoT: The combined quantitative and behavioral evidence characterizes EoT regeneration as reasoning-like continuation into the answering phase, termed spurious CoT termination.

6 Causal Probing of Exit-Token Attention

The paper probes whether insufficient attention to an injected EoT contributes to spurious CoT termination, using EAB across models, benchmarks, and early-exit methods. Increasing attention to the EoT generally reduces spurious termination and answering-phase length, while effects vary by model and sample group.

  • EAB mechanism: EAB adds α to the injected EoT’s pre-softmax attention logit for answering-phase queries, leaving preceding tokens unchanged.Positive α strengthens EoT attention; negative α suppresses it.
  • EAB sweep: Positive α reduces ERR and answering-phase length while largely preserving accuracy, whereas negative α often increases ERR or answer-length inflation.The sweep uses DEER on R1-Distill-14B and Qwen3-14B.
  • Exit-token specificity: Biasing the exit token specifically reduces spurious termination and answering-phase length while preserving accuracy; SoT and Offset-10 show no clear decrease, whereas Offset-1 worsens both behaviors.The comparison covers MATH and GPQA under DEER on R1-Distill-14B.
  • Sample selectivity: EAB selectively helps samples with spurious termination: it reduces ERR below 25% and substantially shortens answers, but can reduce their accuracy.For samples without spurious termination, ERR remains near 0%, accuracy is maintained or improved, and answers become slightly shorter.
  • Generality: Across models and benchmarks, EAB generally reduces ERR and answer length, with especially strong effects for R1-Distill models.For R1-Distill-14B, ERR falls by more than 13 percentage points and answering length nearly halves under DEER; R1-Distill-1.5B also shows roughly halved answering length.
  • Model variation: QwQ-32B shows smaller ERR reductions and sometimes lower accuracy, suggesting some regeneration reflects learned generation patterns that attention intervention does not address.Its No-CoT ERR exceeds 60%, unlike the other models described.
  • Accuracy trade-off: EAB’s accuracy trade-off is larger when early exit removes more reasoning, consistent with truncation of reasoning that would contribute to the final answer.DynaSoR removes more reasoning on average than DEER and correspondingly shows larger accuracy drops.
  • Control baselines: Double-EoT resembles positive EAB, while Block-EoT eliminates regenerated EoT by design but lengthens answers and lowers accuracy without ensuring a clean transition.Prompt-level controls can also mitigate regeneration, but some depend on task-specific boxed-answer formatting.

7 Conclusion

The paper characterizes EoT regeneration as spurious CoT termination: reasoning-like generation continues after forced termination before another EoT appears. Attention intervention supports insufficient attention to the injected EoT as one contributor, while showing that explicit delimiters do not guarantee a successful phase transition.

  • Phenomenon: Forced EoT termination often leads models to generate another EoT later in the answering phase, with prevalence varying across models, benchmarks, and early-exit methods.
  • Behavior: The continuation before regenerated EoT remains reasoning-like, including self-correction markers and boxed-before-EoT patterns, and is termed spurious CoT termination.
  • Mechanism: Increasing attention to the injected EoT reduces regeneration and shortens answering phases, especially for affected samples; token-specific and sample-selective effects support a causal contribution.
  • Conclusion: Injected structural delimiters do not by themselves guarantee a reasoning-to-answering transition because effectiveness depends on subsequent generation attending to and incorporating the EoT.

Limitations

The analysis has methodological and interpretive boundaries involving indirect evidence, deployment of EAB as a practical method, contextual transition signals, and intervention-generation budgets. These constraints qualify how broadly the findings should be interpreted.

  • Evidence: The interpretation of spurious CoT termination relies on indirect behavioral and intervention evidence because the model’s internal state after EoT cannot be directly verified.
  • Practical deployment: Extending EAB from a causal diagnostic to an efficient-reasoning method requires adaptive bias criteria and inference backends that preserve efficient generation while supporting attention modification.
  • Scope of transition signal: The analysis treats EoT as an individual transition signal, although surrounding context and recurring completion patterns may also influence reasoning termination.Boxed-before-EoT patterns and Ans-Prefix effectiveness motivate this broader context as future work.
  • Implementation boundary: Answering-phase generation uses transformers rather than vLLM because attention interventions require per-step attention access, although backend differences are reported as marginal.
  • Generation budget: The intervention pipeline uses the original reasoning budget but a separate answering-phase limit, allowing longer answers than the original budget in intervention runs.Within each α sweep, shared traces and answering budgets preserve comparisons among accuracy, ERR, and length.

A.7 Datasets

The evaluation uses publicly released reasoning models and standard benchmarks, while correctness analyses examine whether EoT regeneration and answer length depend on problem difficulty or exit-point sufficiency.

  • The evaluation covers GSM8K, MATH-500, AMC 2023, AIME 2024, and GPQA-Diamond, with 1,319, 500, 40, 30, and 198 samples, respectively.
  • The study uses publicly released LRMs, datasets, and software libraries, with the benchmarks used for their intended reasoning-evaluation purpose.
  • Full-CoT answer correctness serves as a difficulty proxy, while DEER’s intermediate exit-probe correctness tests whether reasoning at the exit point is sufficient.
  • Correctness does not fully explain longer answering phases: the gap persists among Full-CoT-correct samples, while probe-incorrect cases show higher ERR without consistently longer relative answers.

C.2 Efficiency Analysis

The efficiency analysis compares DEER with and without EAB across runtime and answering-phase behavior, showing that EAB can reduce residual reasoning-like continuation with manageable overhead.

  • Latency measures only answer generation because prefill before the injected EoT is identical across methods, with both vLLM and Transformer implementations reported.
  • EAB preserves comparable throughput when answer lengths are similar because α = 0 adds only a zero-valued bias along the same attention path.
  • EAB can achieve latency comparable to the original vLLM DEER implementation when it substantially shortens answers, while backend differences dominate when token reduction is small.
  • Answer-phase samples containing “Wait” are classified by EoT regeneration and whether non-regenerated cases stop before or at the generation limit.
  • At-limit cases are truncated and have unobservable termination behavior, whereas before-limit non-regenerated cases comprise at most 4.3% of 2,087 evaluation samples.

D.1 Full Results for Answering-phase Length

Across models, methods, and datasets, EoT-regenerated samples have substantially longer answering phases, with larger gaps on harder datasets and larger models.

  • Table 11 compares ERR across exit mechanisms, omitting Natural-EoT cases because they produce no EoT regeneration.
  • EoT-regenerated samples require substantially more answering-phase tokens than samples without regeneration across all evaluated configurations.
  • The answering-phase gap is more pronounced on AIME and GPQA and for R1-Distill-14B and QwQ-32B.
  • Figure 7 provides sample counts with and without EoT regeneration across reasoning-compression-rate bins, complementing per-bin ERR results.

D.3 Pre-regen and Post-regen Analysis

Pre-regen and post-regen analyses characterize spurious CoT termination through token counts, Wait markers, and boxed-expression placement around regenerated EoT tokens.

  • The analysis reports token counts and Wait-token repetition frequencies for R1-Distill-14B, with corresponding results for all models in Table 13.
  • END-BOX5 measures the fraction of samples whose final boxed expression closes within five characters of a segment boundary.
  • The reported metrics distinguish mean boxed-expression count within a segment from the boundary-proximity measure used to identify boxed-before-EoT behavior.
  • For Qwen3-14B and QwQ-32B, boxed-expression density is low near the exit token and peaks immediately before regenerated EoT.
  • In spurious samples, boxed expressions concentrate in the pre-regen span, indicating that boxed answers often appear immediately before regenerated EoT.

D.5 Selectivity of EAB for Samples with Spurious CoT Termination

EAB is evaluated for whether it selectively corrects samples with spurious CoT termination while preserving behavior on unaffected samples. Across models, datasets, and bias strengths, its benefits are strongest for spuriously terminated samples, with accuracy and length effects varying by setting.

  • Selectivity of EAB: EAB is evaluated across models and datasets by separating samples according to spurious termination under the unbiased setting.The grouped analysis uses DEER with α = 2 and α = 4, reporting accuracy, ERR, and answering-phase length.
  • Samples with Spurious CoT Termination: EAB reduces ERR for samples with spurious CoT termination, although accuracy can also degrade.This trend generally extends across models, but its strength varies.
  • Samples without Spurious CoT Termination: For samples without spurious CoT termination, ERR remains close to zero and accuracy is generally preserved under EAB.These results suggest little adverse effect on this group.
  • Extended α Sweep: Positive α improves ERR and answering-phase length only up to around four; at α = 8, generation becomes unstable.Large negative α values largely preserve accuracy but substantially increase ERR and answering-phase length.
  • Target Specificity: Biasing the exit token shows the same qualitative pattern on Qwen3-14B, with the effect specific to that token.The target-position ablation covers MATH and GPQA under DEER.

D.9 Control Baseline Details and Extended Results

Control baselines compare alternative formatting and token interventions with EAB, while extended analyses examine target specificity, answer-phase distributions, and correctness changes. These controls broadly match EAB qualitatively, but sequence-level interventions cannot isolate attention effects and EAB can alter reasoning or expose scoring artifacts.

  • Control Baselines: Double-EoT variants reduce ERR and shorten the answer phase, qualitatively matching positive EAB.The append variant is closer to EAB than the inline variant.
  • Control Baselines: The inline Double-EoT variant has ERR of 6.4% and longer answer phases than the append variant.Append preserves the default boundary pattern, whereas inline stacks two EoT tokens within the boundary.
  • Control Baselines: Double-EoT cannot isolate attention effects because it changes the token sequence and key-value cache, moving attention only indirectly.An extra EoT adds an EoT key that answering-phase queries can attend to.
  • Correctness Changes under EAB: EAB’s case-study outcomes include termination at the generation budget, corrected answers after regeneration, altered reasoning traces, and scoring artifacts.The first three categories reflect changes in model behavior, while the fourth reflects the string-match measurement criterion.
  • Correctness Changes under EAB: Budget-reaching answering phases can either shrink to a few tokens or grow to the limit under EAB, sometimes producing genuine reasoning or repeated “Wait, let me continue” loops.Such cases are often scored incorrect because the answering phase lacks a boxed expression.
  • Correctness Changes under EAB: Some EAB correctness changes arise from reasoning-trace alterations rather than merely facilitating the reasoning-to-answering transition.The paper separately identifies cases where free-form answers are correct but boxed-answer parsing fails.
Loading 2609.03633v1…