- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper introduces “termination misalignment,” a phenomenon in which large reasoning models recognize that their reasoning has stalled or that an input is unsolvable, yet fail to stop reasoning or abstain from answering. It takes two forms: “non-termination” and “specious termination.” The authors develop VB-AF, a gray-box fuzzing framework that generates unsolvable inputs by surrounding a decryption instruction with random gibberish tokens. Across 6,210 inference runs involving 23 models, 73.94% of trajectories exhibit termination misalignment. In 32.77% of these failure trajectories, models explicitly acknowledge stalled progress or unresolvability, or draft termination-appropriate conclusions, yet continue reasoning for thousands of tokens, and model size does not correlate with performance. Adding a few epistemic principles to the system prompt saves more than 5,000 tokens on average and substantially improves correct abstention rates, with little impact on standard benchmark accuracy. A stepwise log-odds analysis shows that models strongly favor continuing to reason throughout their trajectories. The guidance does not weaken this preference; instead, it helps models reach a semantic state conducive to termination earlier. The authors argue that this is a problem of reasoning regulation rather than cognition and highlight VB-AF’s dual-use risks for resource exhaustion and safety bypasses.
Critical strengths
-
VB-AF turns otherwise difficult-to-reproduce reasoning anomalies into a systematic experimental setting, enabling comparisons across 23 models and 6,210 runs. Ablations on instruction wording, vocabulary, delimiters, and insertion position, together with supplementary experiments on naturally unanswerable mathematical problems, strengthen the evidence for the phenomenon and provide a practical stress-testing tool for future research.
-
Beyond examining whether models ultimately abstain, it distinguishes failure to terminate reasoning from failure to abstain after termination, and analyzes cases in which models continue reasoning after explicitly acknowledging unsolvability. The contextual guidance experiments further demonstrate that this gap can be mitigated, establishing “when to stop reasoning” as a control problem worthy of independent study.
Critical weaknesses
-
The experimental setting differs from real deployment scenarios, limiting the generalizability of the headline failure rate. The main experiments rely on random noise and instructions that implicitly suggest the existence of a recoverable message. These differ substantially from situations such as tool failures or code-repair loops in real-world agents. The ablations show considerable sensitivity to instruction wording, so 73.94% should be interpreted as the failure rate under this particular stress test. The experiments on naturally unanswerable mathematical problems provide supplementary evidence but are insufficient to establish broad applicability to real-world agent scenarios.
-
The measurement of “recognizing unsolvability” and the explanation of the termination mechanism remain insufficiently supported. Statements of unsolvability in reasoning traces do not directly establish that a model has formed a stable internal judgment, and the aggregate failure rate does not distinguish cases with and without such recognition. The Δthink analysis covers only a small number of models, and the observed patterns do not establish a causal mechanism in which contextual states trigger termination. The proposed explanation that early-termination trajectories are underrepresented in training also remains untested.
Questions for authors
- How reliable are the annotations of explicit recognition? The figure of 32.77% comes from manual screening assisted by regular-expression keyword matching, but the paper reports neither inter-annotator agreement nor whether a second independent annotator was involved. It also does not report how many of the 1,505 trajectories merely mention unsolvability as one possible hypothesis, rather than conclude that the input is unsolvable and then continue reasoning. Could the authors provide the annotation protocol, agreement metrics, and an estimate of the misclassification rate based on a random sample?
Rebuttal by Authors
We thank you for your review, constructive feedback, and highlighting the strengths of our work.
W1
We agree and would like to point out that we have explicitly qualified the failure rate as confined to the 6,210 traces obtained from our specific VB-AF stress-test (Lines 14,18,48-51, 306-308). It should not be interpreted as a claim of the universal failure rate, which we do not attempt to measure. Our ablations (Appendix A.1, A.2) vary inductive bias of the payload (with neutral P3, P4 included) and test models on naturally unanswerable math problems respectively; they find that the phenomenon persists, albeit its prevalence is affected. In Section 3.1 (Lines 158-161) we motivate the introduction of VB-AF in the observation that relevant user reports are hard to reproduce due to specific interaction histories and environments. To reproduce the failure for systematic controlled analysis, we simulate persistent unresolvability and lack of meaningful progress through VB-AF inputs.
W2.1,Q1: Explicit semantic recognition
We took your feedback as an opportunity to conduct a conservative full-corpus, label-blinded re-annotation of all behavioral failure traces (Strict) using our formalized rubric, along with an independent automated LLM annotation pass (Auto, Gemini 3.8 Flash). During this audit we reconciled run-ID alignment and reverified borderline abstentions updating the overall failure count from 4,592 to 4,666 out of 6,210 runs (75.14%, of which 4,579 overlapped with the initial number (Original) after removing 8 runs reclassified as terminated abstentions). Sending the 16,384-token noisy traces to the LLM directly triggered erratic behavior and safety blocks, so we used truncated traces instead, created using ± 400-char context windows around keyphrase hits identified by a regex similar to the one used to guide manual annotation.
We then made our semantic recognition rubric explicit with the following classes of common recognition patterns (each is a sufficient condition):
Inter-annotator agreement/misclassification:
Condition class breakdown (N=4,666)
Nearly half (49.89%) of all recognized failure traces contain explicit acknowledgements of stalled progress or drafted abstention conclusions (C1∪C2), while 50.11% articulate unresolvability as a hypothesis/explanation (C3 only). There are hence 24.86% (1,544/6,210) terminated abstentions, 21.63% (1,343) failure runs with explicit recognition, and 53.51% (3,323) failure runs without explicit recognition.
W2.2: Logit analysis, causal mechanism, training hypothesis
As noted in Section 4.2,5 (Lines 441-443, 511-512) extracting the full-vocabulary logit distribution over a 32,768 token budget required vLLM V0 execution which constrained models to those that could fit within our compute budget and be representative of our non-Robust clusters. We do not claim to establish a complete mechanism, and the logit analysis serves its purpose to falsify the possible alternate hypothesis that priming may gradually decay step-level continuation log-odds instead of altering the reasoning trajectory to reach a contextual state earlier. We explicitly mention that training under-representation is included as a discussion of a possible mechanism to verify in future work rather than a formal hypothesis that has been proved (Section 4.2 Lines 500-503) in the paper.
We have updated the paper with the new experiment and metric deltas (Section 3.3,3.5; Appendix C.3), made the distinct value of logit analysis clearer (Section 4.2), and updated the AI use disclosure in light of the automated annotation pass.