- Rating
- 2
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper introduces GLOBE, a training-free early stopping method to stop when either the model is ready to answer or reasoning appears to be going in circles. They evaluate Qwen3.6-35B-A3B and Gemma-4-12B-it on five tasks: CSQA, Character-LLM, GSM8K, BFCL tool calling, and agent memory and show GLOBE preserves acc while cutting tokens up to 50%.
Critical strengths
Using geometry as a proxy for detecting redundancy in reasoning is an interesting idea.
Critical weaknesses
-
Tuning on test benchmarks. GLOBE has four main hyperparameters: a confidence threshold, a phase-switch threshold, and two thresholds for geometric stopping. However, Section 4.5 states that the geometric thresholds were selected using the 2,251 Qwen BFCL v4 Live items reported in Table 1. This means that these hyperparameters were tuned on a reported test benchmark, raising concerns about the reliability of that result and the generalization of the selected thresholds.
-
Limited novelty and missing comparisons. GLOBE motivates two phases: one in which the model is confident and already has sufficient information to answer, and another in which it is stuck in redundant or unproductive reasoning. A similar distinction is already considered in [1,2,3]. [1] uses a fixed confidence threshold of 0.95 to stop reasoning when the model is confident. [3] extends confidence-based stopping with an additional criterion for detecting repetitive reasoning, which is closely related to GLOBE's motivation. [2] also considers related stopping criteria. The authors do not compare against these, making it difficult to assess the additional benefit of GLOBE's geometric formulation.
-
Missing evaluations on harder benchmarks. The paper does not evaluate GLOBE on challenging benchmarks such as AIME or LiveCodeBench. These evaluations are important because geometric stopping could be particularly useful on longer, more complex reasoning traces. GSM8K alone provides limited evidence for this setting, as its relatively easy questions may not elicit redundant reasoning.
-
Sensitivity to segmentation and embedding context. A potential limitation is how reasoning traces are divided into segments for embedding. Ideally, these segments should be semantically meaningful, but sentence boundaries may not reliably capture reasoning steps. Moreover, the same sentence can serve different roles at different points in a trace, yet embedding it independently produces the same vector regardless of its surrounding context. This raises questions about whether the resulting geometry accurately captures reasoning progress. Frequent embedding calls also introduce overhead, and hurting efficiency.
[1] Yang, Chenxu, et al. "Dynamic Early Exit in Reasoning Models."
[2] Wang, Xi, et al. "Conformal Thinking: Risk Control for Reasoning on a Compute Budget."
[3] Hosseini, Parsa, et al. "Early Stopping for Large Reasoning Models via Confidence Dynamics."
Questions for authors
Could the authors provide more insight about segment reasoning traces and embed the resulting segments?
Could the authors also evaluate GLOBE on a broader range of models? The current evaluation covers only two models.
Finally, I would be curious to see how GLOBE compares with other baselines on more difficult benchmarks.
4 comments
Response to Reviewer nBvC (2/3): Full 5-Task Main-Table Comparison with CoDE-Stop & Conformal Thinking [W2b]
2/3 — How do CoDE-Stop and Conformal Thinking compare on the five Main-Table tasks? [W2b]
Below is the complete 5-benchmark Gemma comparison, which we will incorporate into the revised Main Table.
Direct Answer: Accounting for full end-to-end serving costs—and even when excluding offline calibration time entirely—GLOBE achieves the fastest E2E on all 5/5 tasks (3,001s mean) with 78.08% 5-task mean score, outperforming CoDE-Stop (77.31%, slower) and running faster online ( faster including calibration) than target-calibrated Conformal Thinking (78.58%).
Because Conformal Thinking's 400–512-item per-task calibration and probing took (31h) on 12B Gemma alone, multi-task calibration on 35B Qwen exceeded the rebuttal window; below is the full 5-task matrix on Gemma-4-12B-it (bold = best among thinking methods including Full CoT; Agent Memory scores QA sessions and sums Tok/E2E across all dialogues over KV + Vector + RecSum). To faithfully reflect real deployment wall time, E2E includes main decoding, all trial/probe requests, continuations, and Conformal's mandatory per-task offline calibration ( unpacked below). Paper Table-1 rows report retained output tokens, whereas CoDE-Stop and Conformal logs record billed generated tokens including trial probes (efficiency comparisons below therefore rely on wall-clock E2E):
Acc = accuracy (%); NAS = Character-LLM's rubric-based Normalized Acting Score; Tok = per-item tokens; E2E = active task seconds; Avg. = 5-task mean.
| Acc | Tok | E2E | | NAS | Tok | E2E | | Acc | Tok | E2E | | Acc | Tok | E2E | | Acc | Tok | E2E | | Acc | Tok | E2E | |
| Nothink | 81.2 | 342 | 837 | 91.7 | 117 | 497 | 94.0 | 160 | 635 | 77.1 | 51 | 888 | 31.8 | 766 | 2,683 | 75.16 | 287 | 1,108 | |||||
| Full CoT | 83.0 | 1,939 | 5,009 | 91.5 | 699 | 9,893 | 94.3 | 1,382 | 4,047 | 77.6 | 601 | 5,782 | 41.9 | 2,308 | 7,523 | 77.66 | 1,386 | 6,451 | |||||
| DEER | 77.9 | 743 | 2,053 | 94.3 | 646 | 2,611 | 88.6 | 1,414 | 4,846 | 76.4 | 526 | 5,600 | 38.7 | 1,951 | 7,286 | 75.18 | 1,056 | 4,479 | |||||
| ThinkBrake | 82.1 | 1,831 | 4,792 | 92.8 | 500 | 7,901 | 95.4 | 1,424 | 4,322 | 76.7 | 386 | 3,693 | 38.3 | 2,217 | 6,671 | 77.06 | 1,272 | 5,476 | |||||
| RCPD | 82.2 | 1,268 | 3,624 | 93.9 | 1,175 | 5,458 | 89.7 | 1,066 | 3,865 | 77.6 | 466 | 4,962 | 42.6 | 2,210 | 7,120 | 77.20 | 1,237 | 5,006 | |||||
| CoDE-Stop | 83.29 | 1,872 | 4,722 | 93.91 | 710 | 2,954 | 92.65 | 1,448 | 4,042 | 77.57 | 447 | 3,966 | 39.14 | 2,992 | 6,321 | 77.31 | 1,494 | 4,401 | |||||
| Conformal | 84.19 | 2,231 | 9,705 | 93.91 | 704 | 27,529 | 94.84 | 1,432 | 7,577 | 77.61 | 835 | 16,633 | 42.37 | 3,210 | 50,145 | 78.58 | 1,682 | 22,318 | |||||
| GLOBE (Ours) | 82.0 | 689 | 2,034 | 92.4 | 645 | 2,478 | 95.1 | 717 | 2,306 | 81.1 | 306 | 2,744 | 39.8 | 1,708 | 5,442 | 78.08 | 813 | 3,001 |
Key takeaways and E2E decomposition:
- Faster on all 5/5 tasks—both with and without offline calibration: Conformal (Wang et al., 2026) reaches 78.58% via 400–512 labeled target items per task (defaulting to 0 stops / 100% Full CoT fallback on Character-LLM and GSM8K), splitting 22,318s into (CSQA , Character , GSM8K , BFCL , Memory ). Even excluding completely, Conformal's pure online (6,507s, 2,996s, 3,866s, 8,828s, 38,624s) and CoDE-Stop (4,401s mean) are slower than GLOBE on every single task ( faster than Conformal online, total)! All methods share identical per-task concurrency ( CSQA/GSM8K, BFCL/Memory, Paper Appendix C.1, Table 3; only Character-LLM ran CoDE-Stop/Conformal at in the rebuttal window vs. in Table 1—which we will unify to in the revision, though GLOBE's 2,478s already beats both).
- GLOBE vs. CoDE-Stop & probing overhead: Without target labels, GLOBE beats CoDE-Stop in 5-task mean (78.08% vs. 77.31%, +0.77 pts; +3.53 BFCL, +2.45 GSM8K, +0.66 Memory) at speed (3,001s vs. 4,401s). Whereas DEER (4,479s), CoDE-Stop (4,401s), and Conformal (12,164s online) keep firing LLM probes across deep traces, GLOBE retires probes at and relies on async segment embeddings (17.63 ms P50 at 80 tokens on main 8B; 8.15 ms on lightweight 0.6B, Comment 3/3).
Response to Reviewer nBvC (3/3): Split Hard-Benchmark Tables, Context Ablation & Single-A100 Perf Latency [W3, W4, Q1, Q3]
3/3 — Hard tasks, context & embeddings [W3, W4, Q1, Q3]
1. Hard tasks: AIME 2026 (), LiveCodeBench v6 (, all new v6 problems; earlier v1–v5 deferred to revision due to time limits), and FRAMES () (1st, ). Qwen3.6-35B-A3B:
| Acc | Tok | E2E | | Acc | Tok | E2E | | Acc | Tok | E2E | | Acc | Tok | E2E | |
| Nothink | 80.00 | 5,246 | 409 | 42.29 | 4,204 | 1,735 | 54.73 | 496 | 997 | 59.01 | 3,315 | 1,047 | |||
| Full CoT | 93.33 | 32,945 | 2,398 | 75.43 | 38,971 | 17,816 | 2,214 | 2,922 | 77.01 | 24,710 | 7,712 | ||||
| DEER | 93.33 | 32,589 | 2,465 | 60.00 | 11,545 | 59.59 | 924 | 70.97 | 18,267 | 5,272 | |||||
| ThinkBrake | 93.33 | 30,431 | 2,486 | 67.43 | 24,102 | 12,355 | 61.77 | 2,208 | 2,979 | 74.18 | 18,914 | 5,940 | |||
| RCPD | 73.33 | 22,533 | 1,716 | 50.86 | 10,285 | 4,453 | 58.50 | 1,672 | 60.90 | 11,276 | 2,614 | ||||
| GLOBE | 93.33 | 25,416 | 62.50 | 1,110 | 1,855 |
Gemma-4-12B-it:
| Acc | Tok | E2E | | Acc | Tok | E2E | | Acc | Tok | E2E | | Acc | Tok | E2E | |
| Nothink | 60.00 | 13,210 | 2,535 | 38.29 | 800 | 641 | 55.34 | 121 | 1,915 | 51.21 | 4,710 | 1,697 | |||
| Full CoT | 66.67 | 37,872 | 6,875 | 71.43 | 15,502 | 13,907 | 59.83 | 4,100 | 11,070 | 65.98 | 19,158 | 10,617 | |||
| DEER | 56.67 | 43,024 | 7,870 | 50.29 | 8,448 | 6,880 | 58.37 | 870 | 3,496 | 55.11 | 17,447 | ||||
| ThinkBrake | 53.33 | 41,331 | 8,714 | 63.43 | 59.83 | 4,177 | 11,447 | 58.86 | 17,739 | 8,768 | |||||
| RCPD | 53.71 | 3,833 | 2,970 | 55.95 | 57.66 | 4,546 | |||||||||
| GLOBE | 21,375 | 3,739 | 12,606 | 9,875 | 58.74 | 2,090 | 5,877 | 12,024 | 6,497 |
Using DEER's official codebase and ThinkBrake's paper hyperparameters, we found both rarely stop on AIME (1–2/30 and 0/30); GLOBE ranks first among early-exit methods in mean Acc on Qwen (75.56%, ) and Gemma (62.98%, ), keeping all 6/6 gaps pp.
2. Context (11 stopped AIME, 9/11 right): sent80 embeds adjacent -char chunks; Expanding Prefix embeds 80 tok, then merges each new 80 tok into the prefix (160, 240, ...) to take angles vs. all prior CoT; Sliding 512-char uses the last 512 chars; Random 0–512-char adds 0–512 random prior chars:
| Flips | Mean | Min | Stops | Can Stop | Uncertain | Continue | |
| sent80 | 3,887 | 0.37 | 0.08 | 11 | 8 | 1 | 2 |
| Expanding Prefix | 3,552 | 0.44 | 0.15 | 0 | - | - | - |
| Sliding 512-char | 2,765 | 0.55 | 0.28 | 0 | - | - | - |
| Random 0–512-char | 3,134 | 0.46 | 0.13 | 1 | 0 | 0 | 1 |
Cumulative/sliding windows dilute turns (, 0 stops); random cuts mid-proof. Trace audits show sent80 halts re-checks once solved: Problem 19 (base-9): 279 found, halting 899/889/898 checks; Problem 2 (palindromes): summed, halting recounts; Problem 12 (spheres): (161) solved, halting re-checks.
3. 4B/0.6B embeddings & Perf:
| Acc | Tok | E2E | | Acc | Tok | E2E | | Acc | Tok | E2E | | Acc | Tok | E2E | |
| 8B | 79.96 | 241 | 3,071 | 95.91 | 358 | 1,751 | 57.99 | 2,040 | 5,730 | 93.33 | 23,174 | 1,767 | |||
| 4B | 79.61 | 236 | 3,081 | 96.36 | 357 | 1,893 | 54.41 | 2,335 | 5,568 | 90.00 | 24,015 | 1,798 | |||
| 0.6B | 79.12 | 243 | 3,177 | 95.91 | 363 | 1,762 | 55.91 | 2,327 | 5,309 | 90.00 | 25,051 | 1,941 |
8B leads thinking methods in mean Acc (81.80%; 2nd in E2E, 3,080s). 0.6B (1.2 GB VRAM) nearly matches 8B (80.24%, +8.43 pp vs. RCPD) and ranks 2nd in E2E (3,047s; – faster than DEER, ThinkBrake, Full CoT). A100 () Perf is 8.15–27.87 ms P50 on 0.6B at 80–320 tok:
| P50 | P90 | P50 | P90 | P50 | P90 | |
| 80 | 17.63 | 24.04 | 19.87 | 23.36 | 8.15 | 9.40 |
| 160 | 34.05 | 35.05 | 17.33 | 34.73 | 9.20 | 12.90 |
| 320 | 42.84 | 44.15 | 32.77 | 35.17 | 27.87 | 31.17 |
| 1,024 | 90.47 | 94.82 | 54.39 | 58.61 | 34.32 | 37.29 |
| 4,096 | 338.69 | 342.02 | 201.08 | 209.73 | 70.85 | 72.03 |
| 8,192 | 707.83 | 723.34 | 452.05 | 469.98 | 144.84 | 147.11 |
We sincerely appreciate your feedback and hope our extensive new experiments and clarifications will encourage you to reconsider your rating.
Dear Reviewer nBvC,
Thank you again for your thoughtful review and valuable suggestions. We particularly appreciate your emphasis on stronger baseline comparisons and more challenging evaluations.
We fully recognize the merits of existing early-stopping approaches. DEER was already included in our original submission, and we have now additionally incorporated CoDE-Stop and Conformal Thinking, which offer valuable alternative perspectives on reducing reasoning computation and balancing efficiency with answer quality.
While these methods have their respective strengths, we believe actual end-to-end latency, including stopping overhead and calibration costs where applicable, is especially important in practical deployment. Under a unified streaming serving setup, GLOBE achieves the lowest E2E latency on all five Gemma tasks, running 1.47× faster than CoDE-Stop and 4.05× faster online than Conformal Thinking, while maintaining competitive accuracy. We also explicitly report the accuracy trade-offs and calibration overhead.
Beyond baseline comparisons, we added DeepSeek-R1-Distill-Qwen-14B, AIME 2026, LiveCodeBench v6, and FRAMES. To clarify the generalization protocol, Qwen BFCL v4 Live was our sole threshold-selection set, chosen for its moderate difficulty and tool-use characteristics. All other models and tasks in the original main table used the same frozen thresholds without retuning, as did these additional evaluations.
These results further distinguish GLOBE's contribution: combining early answer readiness with later trajectory-based progress detection through a training-free phase switch, without target-task calibration or preset stopping budgets.
We understand how demanding the review process can be and sincerely appreciate your time. As the author–reviewer discussion period ends on November 18 (AoE), we would be especially grateful if you could share whether these additional comparisons and evaluations address your concerns, or whether any questions remain that we could clarify before the deadline.
Thank you very much for your consideration.
Best regards,
The Authors
Response to Reviewer nBvC (1/3): Stopping Overhead vs. DEER/CoDE-Stop/Conformal & R1-Distill-14B Transfer [W1, W2a, Q2]
Thank you for your feedback and references [1–3], which made foundational advances in reasoning early stopping—reflection-triggered trial exits ([1] DEER Yang et al., 2026), risk calibration ([2] Conformal Thinking Wang et al., 2026), and confidence dynamics ([3] CoDE-Stop Hosseini et al., 2026). We will cite and add CoDE-Stop and Conformal Thinking to our revised Main Table.
Comments: 1/3 Novelty beyond [1–3] & R1-Distill-14B; 2/3 Gemma Main Table vs. CoDE-Stop & Conformal; 3/3 Hard tasks, context & 8B/4B/0.6B.
GLOBE models spherical reasoning trajectories to unify early trial-answer exits with a later trajectory stop in a training-free online algorithm. At the phase switch, GLOBE stops trial-answer probing and switches to geometry-based stopping; repeated reversals with minimal net progress then trigger a stop. Chosen on Qwen BFCL v4 Live, GLOBE cuts retained tokens by 49.7% and active end-to-end wall time (E2E) by 51.0% vs. Full CoT across five tasks and two models while preserving mean scores.
Prior work so far reports token counts, offline replay, or small-subset timing; at submission time, GLOBE is the first early-stopping work to systematically report live streaming
vLLM + SSEE2E as a primary Main-Table metric—re-engineering all baselines onto the same streaming and prefix-KV stack for fair comparison (summing main decoding, probes/trials incl. discarded work, and continuations; async embeddings overlap decoding, counted once)—and to evaluate long-horizon multi-turn agents (Agent Memory across KV, Vector, and Recursive Sum, 3×192 dialogues: 37 memory-writing + 155 scored QA/backend, interleaving reasoning, tools, observations, and memory updates while closing only<think>).Spanning eight benchmarks (CSQA, Character-LLM, GSM8K, BFCL v4 Live, Agent Memory, AIME 2026, LiveCodeBench v6, FRAMES) and three models (Qwen3.6-35B-A3B, Gemma-4-12B-it, DeepSeek-R1-Distill-Qwen-14B), we will open-source GLOBE and our serving and evaluation code.
1/3 — Novelty & live serving beyond [1–3], plus R1-Distill-14B [W1, W2a, Q2]
Building on [1–3], GLOBE adds two advances: (a) Method: rather than probing across
<think>or calibrating on target labels, GLOBE unifies early trials with training-free spherical trajectory geometry (sent80Kt,Qt) via a one-way phase switch that retires probes at Kt=4 for ~8–18 ms async embeddings; (b) Evaluation: GLOBE benchmarks live streamingvLLM + SSEE2E as a Main-Table metric and long-horizon multi-turn agents (Agent Memory), addressing deployment limits in [1–3]:vLLM + SSEand prefix-KV stack (4,964s) for fair comparison, yet unretired probes still erase deep-trace E2E gains and stop only 1–2/30 AIME items; DTSR (Xiang et al., ACL 2026) likewise shows DEER cuts tokens yet raisesvLLMlatency without trained classifiers.<think>, incurring higher online overhead in our evaluation (1.47× slower and 0.77 mean pts lower than GLOBE; see Comment 2/3).Zero-retuning transfer to DeepSeek-R1-Distill-Qwen-14B: To cover the DeepSeek-R1-Distill family commonly evaluated in [1–3], ThinkBrake, and RCPD, we transfer our frozen Qwen BFCL thresholds (0.90,4,8,0.14) (-48.4% Tok, -52.5% E2E, +0.74 pts across 9 other Table-1 pairs) to DeepSeek-R1-Distill-Qwen-14B. Because its chat template lacks native multi-turn tool schemas in our
vLLMrunner (BFCL, Agent Memory), we run all three non-tool tasks (CSQA, Character-LLM, GSM8K), where GLOBE ranks first on all 3-task means (76.34% vs. 76.30%, 544 Tok [-18.7%], 1,169s [-26.0%]):