acceptodds
Under review as a conference paper at ICLR 2027

GLOBE: Geometry-Led Online Behavioral Early Exit via Unit-Sphere Trajectories

Abstract

Chain-of-thought (CoT) reasoning can improve problem solving, but models often keep generating after the correct answer is already reachable; the extra steps raise serving costs and may replace a correct intermediate result with an incorrect one. We propose GLOBE (Geometry-Led Online Behavioral Early Exit), an online, training-free algorithm that terminates the model's CoT process early. GLOBE embeds each completed reasoning segment as a point on a unit sphere and reads the shape of the resulting semantic trajectory. GLOBE first applies a trial phase that generates short trial answers in a separate branch, stopping the main stream on an accepted trial after checking confidence against a threshold. If generation continues, a trajectory phase stops the stream once this path repeatedly reverses direction with little net progress. On five benchmarks including mathematical reasoning, tool calling, and agent memory, GLOBE cuts output tokens by 49.7% and active end-to-end wall time (E2E) by 51.0% across two models, and achieves a five-task average of 80.58%, comparable to Full CoT.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Reasoning models, test-time scaling, adaptive inference, early exit, chain-of-thought, efficient inference.

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_nBvCOpenReview ↗
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
  1. 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.

  2. 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.

  3. 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.

  4. 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
AuthorsOpenReview ↗

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 + SSE E2E 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, 3×1923\times 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 (sent80 Kt,QtK_t, Q_t) via a one-way phase switch that retires probes at Kt=4K_t = 4 for ~8–18 ms async embeddings; (b) Evaluation: GLOBE benchmarks live streaming vLLM + SSE E2E as a Main-Table metric and long-horizon multi-turn agents (Agent Memory), addressing deployment limits in [1–3]:

  • [1] DEER (Yang et al., 2026): Serial stop-and-probe DEER is 2.56×2.56\times slower than Full CoT on Qwen BFCL (12,516s vs. 4,888s). In Table 1, we re-engineered DEER onto GLOBE's streaming vLLM + SSE and 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 raises vLLM latency without trained classifiers.
  • [2] Conformal Thinking (Wang et al., 2026): Needs 400–512 labeled target items/task (Tcalib=10,153sT_{\mathrm{calib}} = 10,153\text{s}) plus online probes (Tonline=12,164sT_{\mathrm{online}} = 12,164\text{s})—infeasible without target labels (4.05×4.05\times slower online / 7.44×7.44\times total vs. GLOBE, 2/3).
  • [3] CoDE-Stop (Hosseini et al., 2026): Uses offline rule calibration on AMC23 and repeated checks throughout <think>, incurring higher online overhead in our evaluation (1.47×1.47\times 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)(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 vLLM runner (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%]):

Method\bf MethodCSQA\rlap{\bf CSQA}Character-LLM\rlap{\textbf{Character-LLM}}GSM8K\rlap{\bf GSM8K}Avg.\rlap{\bf Avg.}
Acc ⁠Tok ⁠E2E⁠ ⁠NAS ⁠Tok ⁠E2E⁠ ⁠Acc ⁠Tok ⁠E2E⁠ ⁠Acc ⁠Tok ⁠E2E
Nothink72.4012124559.449830087.8716855073.24129365
Full CoT82.236911,39659.405851,54887.267311,79376.306691,579
DEER80.675371,09058.865782,03287.875581,41975.805581,514
ThinkBrake80.596851,38158.525921,60187.797391,81175.636721,598
RCPD79.775481,14433.874241,22284.996811,67066.215511,345
GLOBE (Ours)80.755941,00460.095181,39488.175211,10976.345441,169
AuthorsOpenReview ↗

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%, 1.47×1.47\times slower) and running 4.05×4.05\times faster online (7.44×7.44\times faster including calibration) than target-calibrated Conformal Thinking (78.58%).

Because Conformal Thinking's 400–512-item per-task calibration and probing took >110,000s>110{,}000\text{s} (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 3×1553\times 155 QA sessions and sums Tok/E2E across all 3×1923\times 192 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 (Tcalib+TonlineT_{\mathrm{calib}} + T_{\mathrm{online}} 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.

Method\bf MethodCSQA\rlap{\bf CSQA}Character-LLM\rlap{\textbf{Character-LLM}}GSM8K\rlap{\bf GSM8K}BFCL v4 Live\rlap{\textbf{BFCL v4 Live}}Agent Memory\rlap{\textbf{Agent Memory}}Avg.\rlap{\bf Avg.}
Acc ⁠Tok ⁠E2E⁠ ⁠NAS ⁠Tok ⁠E2E⁠ ⁠Acc ⁠Tok ⁠E2E⁠ ⁠Acc ⁠Tok ⁠E2E⁠ ⁠Acc ⁠Tok ⁠E2E⁠ ⁠Acc ⁠Tok ⁠E2E
Nothink81.234283791.711749794.016063577.15188831.87662,68375.162871,108
Full CoT83.01,9395,00991.56999,89394.31,3824,04777.66015,78241.92,3087,52377.661,3866,451
DEER77.97432,05394.36462,61188.61,4144,84676.45265,60038.71,9517,28675.181,0564,479
ThinkBrake82.11,8314,79292.85007,90195.41,4244,32276.73863,69338.32,2176,67177.061,2725,476
RCPD82.21,2683,62493.91,1755,45889.71,0663,86577.64664,96242.62,2107,12077.201,2375,006
CoDE-Stop83.291,8724,72293.917102,95492.651,4484,04277.574473,96639.142,9926,32177.311,4944,401
Conformal84.192,2319,70593.9170427,52994.841,4327,57777.6183516,63342.373,21050,14578.581,68222,318
GLOBE (Ours)82.06892,03492.46452,47895.17172,30681.13062,74439.81,7085,44278.088133,001

Key takeaways and E2E decomposition:

  1. 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 Tcalib=10,153s+Tonline=12,164sT_{\mathrm{calib}} = 10{,}153\text{s} + T_{\mathrm{online}} = 12{,}164\text{s} (CSQA 3,197+6,5073{,}197+6{,}507, Character 24,533+2,99624{,}533+2{,}996, GSM8K 3,711+3,8663{,}711+3{,}866, BFCL 7,805+8,8287{,}805+8{,}828, Memory 11,521+38,62411{,}521+38{,}624). Even excluding TcalibT_{\mathrm{calib}} completely, Conformal's pure online TonlineT_{\mathrm{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 (4.05×4.05\times faster than Conformal online, 7.44×7.44\times total)! All methods share identical per-task concurrency (c=8c=8 CSQA/GSM8K, c=4c=4 BFCL/Memory, Paper Appendix C.1, Table 3; only Character-LLM ran CoDE-Stop/Conformal at c=4c=4 in the rebuttal window vs. c=1c=1 in Table 1—which we will unify to c=1c=1 in the revision, though GLOBE's c=1c=1 2,478s already beats both).
  2. 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 1.47×1.47\times 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 Kt=4K_t = 4 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).
AuthorsOpenReview ↗

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 (N=30N=30), LiveCodeBench v6 (N=175N=175, all new v6 problems; earlier v1–v5 deferred to revision due to time limits), and FRAMES (N=824N=824) (1st, 2nd‾\underline{2\text{nd}}). Qwen3.6-35B-A3B:

Method\bf MethodAIME 2026\rlap{\textbf{AIME 2026}}LiveCodeBench v6\rlap{\textbf{LiveCodeBench v6}}FRAMES\rlap{\bf FRAMES}Avg.\rlap{\bf Avg.}
AccTok ⁠E2E⁠ ⁠AccTok ⁠E2E⁠ ⁠AccTok ⁠E2E⁠ ⁠AccTok ⁠E2E
Nothink80.005,24640942.294,2041,73554.7349699759.013,3151,047
Full CoT93.3332,9452,39875.4338,97117,81662.26‾\underline{62.26}2,2142,92277.0124,7107,712
DEER93.3332,5892,46560.0021,287‾\underline{21,287}11,54559.599241,807‾\underline{1,807}70.9718,2675,272
ThinkBrake93.3330,4312,48667.4324,10212,35561.772,2082,97974.1818,9145,940
RCPD73.3322,5331,71650.8610,2854,45358.501,009‾\underline{1,009}1,67260.9011,2762,614
GLOBE93.3323,174‾\underline{23,174}1,767‾\underline{1,767}70.86‾\underline{70.86}25,41611,333‾\underline{11,333}62.501,1101,85575.56‾\underline{75.56}16,567‾\underline{16,567}4,985‾\underline{4,985}

Gemma-4-12B-it:

Method\bf MethodAIME 2026\rlap{\textbf{AIME 2026}}LiveCodeBench v6\rlap{\textbf{LiveCodeBench v6}}FRAMES\rlap{\bf FRAMES}Avg.\rlap{\bf Avg.}
AccTok ⁠E2E⁠ ⁠AccTok ⁠E2E⁠ ⁠AccTok ⁠E2E⁠ ⁠AccTok ⁠E2E
Nothink60.0013,2102,53538.2980064155.341211,91551.214,7101,697
Full CoT66.6737,8726,87571.4315,50213,90759.834,10011,07065.9819,15810,617
DEER56.6743,0247,87050.298,4486,88058.378703,49655.1117,4476,082‾\underline{6,082}
ThinkBrake53.3341,3318,71463.437,709‾\underline{7,709}6,142‾\underline{6,142}59.834,17711,44758.8617,7398,768
RCPD63.33‾\underline{63.33}33,235‾\underline{33,235}6,113‾\underline{6,113}53.713,8332,97055.951,345‾\underline{1,345}4,554‾\underline{4,554}57.6612,804‾\underline{12,804}4,546
GLOBE63.33‾\underline{63.33}21,3753,73966.86‾\underline{66.86}12,6069,87558.742,0905,87762.98‾\underline{62.98}12,0246,497

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%, 1.55×1.55\times) and Gemma (62.98%, 1.63×1.63\times), keeping all 6/6 gaps ≤5\le 5 pp.

2. Context (11 stopped AIME, 9/11 right): sent80 embeds adjacent ≥80\ge 80-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:

Context\bf ContextTrajectory\bf TrajectoryES\bf ESAudit (gpt6-astra)\textbf{Audit (gpt6-astra)}
FlipsMean QQMin QQStopsCan StopUncertainContinue
sent803,8870.370.0811812
Expanding Prefix3,5520.440.150---
Sliding 512-char2,7650.550.280---
Random 0–512-char3,1340.460.131001

Cumulative/sliding windows dilute turns (min⁡Q≥0.15\min Q \ge 0.15, 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): 5+15+20+15+6+1=625+15+20+15+6+1=62 summed, halting recounts; Problem 12 (spheres): r=122/39r=122/39 (161) solved, halting re-checks.

3. 4B/0.6B embeddings & c=1c=1 Perf:

Embedding Model\bf Embedding\ ModelBFCL v4 Live\rlap{\textbf{BFCL v4 Live}}GSM8K\rlap{\bf GSM8K}Agent Memory\rlap{\textbf{Agent Memory}}AIME 2026\rlap{\textbf{AIME 2026}}
Acc ⁠Tok ⁠E2E⁠ ⁠Acc ⁠Tok ⁠E2E⁠ ⁠Acc ⁠Tok ⁠E2E⁠ ⁠Acc ⁠Tok ⁠E2E
8B79.962413,07195.913581,75157.992,0405,73093.3323,1741,767
4B79.612363,08196.363571,89354.412,3355,56890.0024,0151,798
0.6B79.122433,17795.913631,76255.912,3275,30990.0025,0511,941

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; 2.0×2.0\times–2.2×2.2\times faster than DEER, ThinkBrake, Full CoT). A100 (c=1c=1) Perf is 8.15–27.87 ms P50 on 0.6B at 80–320 tok:

Tokens\bf Tokens8B\bf 8B4B\bf 4B0.6B\bf 0.6B
P50P90P50P90P50P90
8017.6324.0419.8723.368.159.40
16034.0535.0517.3334.739.2012.90
32042.8444.1532.7735.1727.8731.17
1,02490.4794.8254.3958.6134.3237.29
4,096338.69342.02201.08209.7370.8572.03
8,192707.83723.34452.05469.98144.84147.11

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

Reviewer_E2nwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The study presents GLOBE (Geometry-Led Online Behavioral Early Exit) which is a training-free algorithm that terminates the model’s CoT process early. It embeds full sentences of CoT sis_i using an embedding model in a unit sphere using normalization and stops the CoT generation process if the displacement between sis_i and si+2s_{i+2} results in reverse direction or if the displacement path of the trajectory in embedded space oscilates too much. The study also includes other signals like wait, actually to determine if the CoT should be stopped early.

Critical strengths
  1. The paper provides a way to map the CoT sentences to a unit sphere, and based on the displacement of trajectories of the CoT, it decides whether the chain should be stopped or not.
Critical weaknesses
  1. The paper has no way of telling if a reasoning has been completed and should be truncated as it relies on some signals in vector space. The embeddings are dependent on the embedding model, and the paper does not examine the geometry of this embedding models they are known to be anisotropic. This might not generalize to other embedding models, and the authors ran the experiment using only one embedding model.
  2. There is a possibility that the datasets used in this study are included to train the models used in this study. For the Qwen model on GSM8K and CSQA, Full CoT improves over no-think by <1 point, so these benchmarks cannot test whether the stopping rule preserves necessary reasoning. The accuracy for GSM8k is already 96.7%, but Table 7 says GLOBE Rescue 29.6% and Harm only 0.94% of the items that CoT got wrong, but the reported GLOBE's accuracy is 96.5. This discrepancy should be discussed clearly. Qwen also hurts 10% for agent memory, and it is the task where reasoning is definitely needed as compared to others.
Questions for authors
  1. What is the length of answers that were rescued by the model using GLOBE but were not solved by CoT?
  2. How is KoscK_{osc} defined?
  3. In ablation study, the authors mention that the phase switch costs accuracy, but the average accuracy is worse without the trajectory stop. Please clarify why this is said in the paper.
4 comments
AuthorsOpenReview ↗

Response to Reviewer E2nw (1/3): Stopping-Rule Notation & Stopped vs. Continuing Trace Audit [W1, Q2]

Thank you for examining whether our rule separates repetition from progress, transfers across embedding models, and preserves needed reasoning.

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.

Fewer tokens need not mean lower latency. 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 + SSE E2E 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, 3×1923\times 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.

Comments: 1/3 Rule notation, stopped/continuing trace audits & τQ\tau_Q ablations; 2/3 8B / 4B / 0.6B scaling, cross-family agreement & A100 latency; 3/3 Harder tasks, Table 7 & §4.3.


1/3 — What separates repetitive oscillation from useful progress? [W1, Q2]

GLOBE stops <think> online via trial confidence (Phase 1) and trajectory geometry (Phase 2). While KtK_t and QtQ_t are operational proxies rather than formal completion guarantees, trace audits and τQ\tau_Q ablations confirm that Qt≤0.14Q_t \le 0.14 captures stagnation while Qt>0.14Q_t > 0.14 protects active progress.

1. Rule (Paper §3.3–3.4, Eqs. 2–4; unified notation added to §3.3): Unit segment embeddings uj∈Sd−1u_j \in \mathbb{S}^{d-1} define chord steps ξj=uj+1−uj\xi_j = u_{j+1} - u_j. Reversals KJ=∑j=1J−2I[κj<0]K_J = \sum_{j=1}^{J-2} \mathbb{I}[\kappa_j < 0] (κj=⟨ξj,ξj+1⟩\kappa_j = \langle \xi_j, \xi_{j+1} \rangle) count turns >90∘> 90^\circ. Ratio QJ=∥uJ−u1∥2/∑j=1J−1∥ξj∥2Q_J = \lVert u_J - u_1 \rVert_2 / \sum_{j=1}^{J-1} \lVert \xi_j \rVert_2 measures net progress over path length (Kt=KJ(t),Qt=QJ(t)K_t = K_{J(t)}, Q_t = Q_{J(t)}). Phase 1 (Kt<Kphase=4K_t < K_{\mathrm{phase}} = 4) exits on confidence q>τconf=0.90q > \tau_{\mathrm{conf}} = 0.90. Phase 2 retires probes and stops iff Kt≥Kosc=8K_t \ge K_{\mathrm{osc}} = 8 and Qt≤τQ=0.14Q_t \le \tau_Q = 0.14.

2. Phase-exit audit and Paper Table 2's τQ\tau_Q ablation: Items = problems or scored QA sessions (3×155=4653\times 155 = 465 for Agent Memory); Phase-1/Phase-2 Exits = confidence/geometry stops (Agent Memory final turns). Last column: Paper Table 2's whole-run ablation (without τQ\tau_Q minus GLOBE in pts before rounding; negative = drop without τQ\tau_Q).

Model and Task\bf Model\ and\ TaskNew Stop Audit\bf New\ Stop\ AuditGate Ablation\bf Gate\ Ablation
ItemsPhase-1 ExitsPhase-2 ExitsAcc Change No τQ\tau_Q
Qwen BFCL v4 Live2,2518281,053-0.9
Gemma BFCL v4 Live2,2511,175307-5.1
Qwen Agent Memory46514079-2.3
Gemma Agent Memory4656543-3.4
Qwen GSM8K1,3192871,031-0.3
Gemma GSM8K1,319975320-4.6
  • Continuing traces (Qt>0.14Q_t > 0.14): At first Kt=8K_t = 8, 74.8% (803/1,073) of Qwen BFCL and 86.8% (171/197) of Qwen Agent Memory traces satisfy Qt>0.14Q_t > 0.14 and continue; to validate this, we analyzed these unstopped chains with gpt6-astra and confirmed that they indeed still needed to continue reasoning (e.g., refining tool arguments or memory queries). Removing τQ\tau_Q drops accuracy across all 6 Table-2 pairs (-0.9 Qwen BFCL, -2.3 Qwen Memory, -0.3 Qwen GSM8K, -3.4 to -5.1 pts on Gemma) and lowers Gemma Character-LLM NAS by -2.05 pts (p=8.5×10−4p = 8.5\times 10^{-4}).
  • Stopped traces (Qt≤0.14Q_t \le 0.14): Analyzing the stop points with gpt6-astra confirms that 82.19% (263/320) of Gemma GSM8K and 78.48% (62/79) of Qwen Agent Memory stops had already reached an answer or stalled. Across 502 audited traces (320 Gemma GSM8K, 79 Qwen Memory, 57 Qwen GSM8K, 31 Gemma Memory, 15 Gemma BFCL-100), only 1/502 (0.20%) shows right/wrong switching; closing only <think> preserves content completion.
AuthorsOpenReview ↗

Response to Reviewer E2nw (2/3): Multi-Scale & Cross-Family Embedding Models and A100 Latency [W1, Q2]

2/3 — Does changing the embedding model hurt quality or erase latency savings? [W1, Q2]

Below we evaluate (1) smaller embedding models (8B / 4B / 0.6B), (2) cross-family stopping agreement, and (3) single-A100 (c=1c=1) Perf latency.

Direct Answer: Our original submission already evaluated three embedding models (8B, 4B, and 0.6B) on BFCL v4 Live and GSM8K (Paper Appendix D.9, Table 11). In this rebuttal, we extend the evaluation to two additional benchmarks, Agent Memory and AIME 2026, as shown in the table below.

Scaling down 13.3×13.3\times from 8B to 0.6B (1.2 GB VRAM) achieves nearly lossless accuracy (80.24% vs. 81.80% on 8B), while ranking second in four-task mean E2E (3,047s, 2.0×2.0\times–2.2×2.2\times faster than DEER, ThinkBrake, and Full CoT). We further observe 8.15–27.87 ms P50 embedding latency at 80–320 tokens and 89.3%–98.8% cross-family stopping-decision agreement across BGE, E5, and MPNet.

1. Smaller embedding models (8B / 4B / 0.6B) and cross-family generalization:

  • Scaling down to 4B and 0.6B (Qwen3.6-35B-A3B): To examine how smaller embedding models perform in deployment (0.6B uses only 1.2 GB VRAM), we extend Paper Appendix D.9 (Table 11: BFCL v4 Live and GSM8K) to Agent Memory (3×1553\times 155 scored QA, 3×1923\times 192 dialogues) and AIME 2026 (N=30N=30, c=4c=4):
Embedding Model\bf Embedding\ ModelBFCL v4 Live\rlap{\textbf{BFCL v4 Live}}GSM8K\rlap{\bf GSM8K}Agent Memory\rlap{\textbf{Agent Memory}}AIME 2026\rlap{\textbf{AIME 2026}}
Acc ⁠Tok ⁠E2E⁠ ⁠Acc ⁠Tok ⁠E2E⁠ ⁠Acc ⁠Tok ⁠E2E⁠ ⁠Acc ⁠Tok ⁠E2E
8B79.96240.93,07195.91358.31,75157.992,0405,73093.3323,174.041,767
4B79.61236.23,08196.36357.31,89354.412,335.475,567.5490.0024,015.131,798
0.6B79.12242.53,17795.91362.71,76255.912,327.315,308.8190.0025,051.341,941

Comparable Performance with Smaller Embeddings. Scaling down 13.3×13.3\times from 8B to 0.6B (1.2 GB VRAM) achieves comparable accuracy (80.24% vs. 81.80%, within 1.56 pp), with only a 0.84 pp difference on BFCL and identical 95.91% on GSM8K. The 0.6B model also closely matches 8B in retained reasoning tokens (<1.3%<1.3\% difference on BFCL/GSM8K) and 4-task mean E2E (3,047s vs. 3,080s).

Competitive Performance Against Baselines. Compared against all thinking baselines (Full CoT, DEER, ThinkBrake, and RCPD), 8B ranks first in 4-task mean accuracy (81.80% vs. Full CoT 81.58%), while 0.6B outperforms RCPD (71.81%) by +8.43 pp. Both rank second in 4-task mean E2E behind RCPD (2,562s), running 2.0×2.0\times–2.2×2.2\times faster than DEER (6,103s), ThinkBrake (6,644s), and Full CoT (6,767s), with 0.6B reducing total output tokens by 28%–36% on Memory and AIME.

  • Cross-family generalization beyond Qwen: Across five external models from three families (BAAI/bge-small-en-v1.5, bge-m3, intfloat/e5-small-v2, e5-large-v2, all-mpnet-base-v2), any constant mean shift after normalization cancels in chord steps ξj=uj+1−uj\xi_j = u_{j+1} - u_j (0/360 flips). Despite wide directional-concentration spread (top-1 singular-value energy from 16.9% on MPNet to 76.7% on E5-large-v2), frozen (Kosc=8,τQ=0.14)(K_{\mathrm{osc}}=8, \tau_Q=0.14) closely matches Qwen3-Embedding-8B (Agreement = % of stop/continue decisions matching the 8B embedding model on the same reasoning prefixes, with stopping thresholds unchanged):
EmbeddingModel(Backbone)\bf Embedding\\ Model\\ (Backbone)Model Spec\rlap{\bf Model\\ Spec}Stop/Continue Agreement\rlap{\bf Stop/Continue\\ Agreement}
Params (M)DimEval PrefixesAgreement (%)
BAAI bge-small-en-v1.5 (BERT)3338486198.8
intfloat e5-small-v2 (E5-BERT)3338486198.6
all-mpnet-base-v2 (MPNet)10976886198.5
intfloat e5-large-v2 (E5-BERT)3351,0242892.9
BAAI bge-m3 (XLM-RoBERTa)5681,0242889.3

2. Single-request A100 (c=1c=1) Perf microbenchmark (80–8,192 tokens): Perf measures isolated per-call embedding latency (P50/P90: median and 90th-percentile ms). Local sent80 segments (80–320 tokens, GLOBE's typical input range) take 8.15–27.87 ms P50 on 0.6B (17.63–42.84 ms on 8B), fitting 512-token BERT/E5 limits; rows 1,024–8,192 contrast cumulative-prefix encoding (707.83 ms on 8B), which GLOBE avoids by embedding individual segments rather than repeatedly encoding the full reasoning prefix:

Input Tokens\bf Input\ TokensEmbedding-8B (ms)\textbf{Embedding-8B (ms)}Embedding-4B (ms)\textbf{Embedding-4B (ms)}Embedding-0.6B (ms)\textbf{Embedding-0.6B (ms)}
P50P90P50P90P50P90
8017.6324.0419.8723.368.159.40
16034.0535.0517.3334.739.2012.90
32042.8444.1532.7735.1727.8731.17
1,02490.4794.8254.3958.6134.3237.29
4,096338.69342.02201.08209.7370.8572.03
8,192707.83723.34452.05469.98144.84147.11
AuthorsOpenReview ↗

Response to Reviewer E2nw (3/3): Indispensable-Reasoning Benchmarks, Paper Table 7 & Section 4.3 [W2, Q1, Q3]

3/3 — Does early exit preserve reasoning when reasoning is indispensable? [W2, Q1, Q3]

Below we cover reasoning-intensive benchmarks, Paper Appendix Table 7 Rescue/Harm denominators, and §4.3 under the frozen Qwen BFCL v4 Live thresholds (full AIME 2026 / LiveCodeBench v6 / FRAMES tables: Reviewer nBvC 3/3; R1-Distill-14B: Reviewer nBvC 1/3; CoDE-Stop & Conformal: Reviewer nBvC 2/3).

Direct Answer: On tasks where CoT adds +10.1 to +33.1 pp over Nothink (Agent Memory, AIME 2026, LiveCodeBench v6, FRAMES), GLOBE achieves a favorable accuracy–efficiency trade-off—staying within 5 pp of Full CoT on every task (matching or exceeding Full CoT on Qwen Agent Memory, Qwen AIME 2026, and Qwen FRAMES) while cutting active E2E by 26%–65%.

1. Reasoning-Intensive and Long-Horizon Benchmarks:

  • Long-horizon multi-turn agents (Agent Memory): Unlike static GSM8K/CSQA, where CoT improves Qwen3.6-35B-A3B by <1<1 pt, Agent Memory requires multi-step reasoning, retrieval, and refinement (124/155 Vector and 125/155 KV sessions). Across 3×1923\times192 dialogues (37 memory-writing + 155 scored QA per backend), CoT improves accuracy by +20.5 pts on Qwen (34.5% to 55.0%) and +10.1 pts on Gemma (31.8% to 41.9%).

    GLOBE's performance: In Paper Table 1, GLOBE ranks first overall on Qwen Agent Memory (58.0%, +3.0 pts over Full CoT, -64.9% E2E) and Gemma BFCL Live (81.1%), and first among early-exit methods on Qwen BFCL Live (80.0%). On Gemma Agent Memory, it stays within 2.1 pp of Full CoT (39.8% vs. 41.9%) while achieving the lowest tokens (1,708) and fastest E2E (5,442s).

  • Indispensable Mathematical Reasoning benchmark:AIME 2026 (N=30N=30, 81,920 cap, c=4c=4; Acc %, retained Tok, active E2E s; bold = 1st, underline‾\underline{\text{underline}} = 2nd among thinking methods):

Method\bf MethodAIME 2026(Qwen3.6-35B-A3B)\mathrlap{\substack{\textstyle\textbf{AIME 2026}\cr \textstyle\textbf{(Qwen3.6-35B-A3B)}}}AIME 2026(Gemma-4-12B-it)\mathrlap{\substack{\textstyle\textbf{AIME 2026}\cr \textstyle\textbf{(Gemma-4-12B-it)}}}
Acc ⁠Tok ⁠E2E⁠ ⁠Acc ⁠Tok ⁠E2E
Nothink80.005,24640960.0013,2102,535
Full CoT93.3332,9452,39866.6737,8726,875
DEER93.3332,5892,46556.6743,0247,870
ThinkBrake93.3330,4312,48653.3341,3318,714
RCPD73.3322,5331,71663.33‾\underline{63.33}33,235‾\underline{33,235}6,113‾\underline{6,113}
GLOBE (Ours)93.3323,174‾\underline{23,174}1,767‾\underline{1,767}63.33‾\underline{63.33}21,3753,739

Even with DEER's official codebase and ThinkBrake's paper settings, both show very low AIME trigger rates (1–2/30 and 0/30), saving no time (2,465s / 2,486s vs. 2,398s), while RCPD drops -20.0 pp on Qwen (73.33%)—highlighting GLOBE's unified-parameter transfer. GLOBE matches 93.33% on Qwen (1.36×1.36\times faster, -26.3% E2E) and reaches 63.33% on Gemma (-3.34 pp vs. 66.67%, 1.84×1.84\times faster, -43.6% tokens, -45.6% E2E). Across AIME 2026, LiveCodeBench v6, and FRAMES (Reviewer nBvC 3/3), GLOBE ranks first among all early-exit methods in 3-task mean accuracy on Qwen (75.56%) and Gemma (62.98%), all 6/6 gaps ≤4.57\le 4.57 pp vs. Full CoT.


2. Paper Appendix Table 7 conditional denominators and rescued lengths:

  • Conditional pools & raw counts: As noted in Paper Appendix Table 7's caption, Rescue/Harm characterizes conditional transitions on separate item-paired analysis runs (distinct from Table 1's aggregate serving runs, rather than an item-level decomposition of Table 1): only Full-CoT-correct items can be harmed (P(wrong∣Full right)P(\text{wrong} \mid \text{Full right})), and only Full-CoT-wrong items rescued (P(right∣Full wrong)P(\text{right} \mid \text{Full wrong})). On Qwen GSM8K, the Full-CoT-correct pool (1,272) is 27×27\times larger than the wrong pool (47), so 29.8% Rescue (14/47 recovered; cited as 29.6%) and 0.94% Harm (12/1,272 lost) yield near-equal raw counts (+14 vs. -12). On Qwen Agent Memory (465 scored QA sessions), 24.9% Rescue (51/205) exceeds 10.8% Harm (28/260) in both rate and raw count (+51 vs. -28 = +23 net sessions) (Rescue also exceeds Harm on Gemma GSM8K [40.0% vs. 3.14%] and BFCL [9.75% vs. 4.14%]). Raw n/Nn/N counts will be added to Table 7.
  • Rescued lengths: GSM8K rescues retain 659+146659 + 146 (Problem 100) and 159+49159 + 49 tokens (Problem 321) vs. 0 content on capped Full CoT (Problem 326); on Qwen MATH-450 (20,480 cap), GLOBE rescues 56/75 zero-content items at 1,326.8 mean tokens.

3. Paper §4.3 ("Removing the phase switch costs accuracy"): "Removing" means ablating (disabling) the phase switch (-2.9 Qwen / -4.5 Gemma pts in Table 2 from unretired probes), whereas ablating the Phase-2 trajectory stop hurts more (-3.8 / -11.3 pts; 471 capped Gemma GSM8K failures, Appendix D.5). We will change "Removing" to "Ablating" in §4.3.

We sincerely appreciate your feedback and hope our extensive new experiments and clarifications will encourage you to reconsider your rating.

Dear Reviewer E2nw,

Thank you again for your thoughtful review and constructive suggestions. We sincerely appreciate your concerns regarding geometric reliability, embedding dependence, and premature stopping, which helped us identify several important directions for additional evaluation.

In response, we substantially expanded our experiments with orthogonality controls, a 502-trace stopping audit, and explicit Rescue/Harm analyses, directly examining whether GLOBE incorrectly interrupts productive reasoning. We also tested smaller and alternative embedding models: a 0.6B encoder maintains a favorable accuracy–efficiency trade-off, with 89.3%–98.8% cross-family agreement in stopping decisions.

We further evaluated AIME 2026, LiveCodeBench v6, and FRAMES, including additional early-stopping baselines. Importantly, our original main-table experiments already used one frozen threshold configuration selected solely on Qwen BFCL v4 Live, chosen as a moderately challenging tool-use benchmark. No other task or reasoning model was retuned, and the newly added benchmarks follow this same protocol, providing further evidence of generalization beyond the original evaluation.

We realize that examining these additional results requires considerable time, and we greatly appreciate your consideration. With the author–reviewer discussion closing on November 18 (AoE), we would be especially grateful to hear whether these findings address your main concerns, or if there are remaining questions we could clarify before the deadline.

Thank you again for your valuable feedback.

Best regards,
The Authors

Reviewer_6pUUOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes GLOBE, a training-free method for stopping LLM reasoning early:

  • Early stage—trial-answer confidence: Stop reasoning if the model can already produce a high-confidence answer.

  • Later stage—semantic trajectories: Map reasoning segments onto a unit sphere. Stop when the trajectory repeatedly changes direction with little net displacement, suggesting repetitive reasoning.

Across two models and five tasks, GLOBE reduces output tokens by approximately 49.7% and execution time by 51.0%, while maintaining comparable average performance. The main concern is whether these geometric features reliably indicate reasoning stagnation.

Critical strengths
  1. Substantial efficiency gains while maintaining performance comparable to full reasoning. Across two models and five tasks, GLOBE reduces output tokens by approximately 49.7% and end-to-end time by 51.0%, while achieving comparable average performance. These results demonstrate practical deployment value.

  2. An insightful stopping strategy that combines answer readiness with reasoning progress. GLOBE uses trial-answer confidence early in reasoning and semantic trajectory signals later to detect repetitive reasoning. Ablation results support the complementary roles of these signals and the phase switch, offering a promising training-free approach to dynamically allocating inference computation.

Critical weaknesses
  1. Insufficient validation that the core geometric metrics reflect reasoning stagnation. Directional reversals and a low net-displacement ratio can also occur in trajectories that continue introducing new content. For example, mutually orthogonal segment vectors can satisfy the stopping conditions without repetition. Without length-matched random stopping and fixed-budget controls, the experiments cannot establish whether the gains come from detecting semantic stagnation. This undermines confidence in the paper’s central contribution.

  2. Threshold selection uses evaluation data, and accuracy claims lack uncertainty estimates. The paper selects thresholds and reports final results on the same Qwen BFCL Live examples, potentially yielding optimistic estimates. Moreover, the average score improvement over full reasoning is only 0.54 percentage points, without confidence intervals or repeated-run results. Consequently, the evidence remains insufficient to establish that GLOBE consistently preserves answer quality on unseen data.

Questions for authors
  1. Clarification of the core metrics’ validity: Mutually orthogonal, non-repetitive segment vectors can satisfy both the reversal-count and net-displacement-ratio stopping conditions. How does GLOBE distinguish such cases from genuine semantic stagnation? Do you have analyses controlling for reasoning length, or comparisons with length-matched random stopping and fixed-budget strategies, demonstrating that the geometric signals provide additional predictive value?

  2. Clarification of evaluation independence and result stability: Section 4.5 states that thresholds were selected using the Qwen BFCL Live examples reported in Table 1. Do you have evaluation results on completely independent data that did not inform parameter selection? On those data, what are the confidence intervals or repeated-run results for accuracy differences, and do they support the claim of preserving full-reasoning quality?

6 comments
AuthorsOpenReview ↗

Response to Reviewer 6pUU (1/3): Orthogonality Bound & Cross-Dataset K_osc Trajectory Audit [W1, Q1]

Thank you for your questions on our geometric stopping rule, length controls, and threshold transfer.

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, probes retire; 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.

Fewer tokens need not mean lower latency. 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 + SSE E2E 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, 3×1923\times 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.

Comments: 1/3 Pairwise orthogonality & stop-trajectory audits; 2/3 Fixed/random length cutoffs & unified 5-task R+CR + C splits; 3/3 Transfer to the 9 other Table-1 pairs, R1-Distill-14B, and 3 hard tasks with paired 95% CIs.


1/3 — How does GLOBE separate orthogonal exploration from stagnation? [W1, Q1]

Direct Answer: GLOBE pairs chordal reversals KtK_t with net-displacement ratio QtQ_t to detect local circling rather than directional novelty. An exact 90∘90^\circ chord turn adds 00 reversals, and Qt>0.14Q_t > 0.14 vetoes premature cuts.

1. Pairwise orthogonality (J≥10J \ge 10, Paper Appendix B.2):

  • Local turns: For unit segment vectors uj∈Sd−1u_j \in \mathbb{S}^{d-1}, chord steps ξj=uj+1−uj\xi_j = u_{j+1} - u_j count a reversal only when ⟨ξj,ξj+1⟩<0\langle \xi_j, \xi_{j+1} \rangle < 0. Of 26/751 near-90∘90^\circ BFCL turns (3.46%), 19 move outward and only 7 move inward.
  • Pairwise-orthogonal sequence: With JJ unit vectors at pairwise cosine ρ=0\rho = 0, adjacent chords turn 120∘120^\circ (⟨ξj,ξj+1⟩=−1\langle \xi_j, \xi_{j+1} \rangle = -1), giving KJ=J−2K_J = J - 2 and QJ=1/(J−1)Q_J = 1/(J - 1). Satisfying (8,0.14)(8, 0.14) requires J≥10J \ge 10 (K10=8,Q10=1/9K_{10} = 8, Q_{10} = 1/9). Near-orthogonal pairs/triples (∣cos⁡θ∣≤0.1\lvert \cos\theta \rvert \le 0.1) account for 0.00%–2.71% / 0.00%–0.05% of audited segments; 10 consecutive orthogonal segments occur in 0.00%.

2. Trajectory-stop audit across five model–task pairs: Items = problems or scored QA sessions (3×155=4653\times155=465 for Agent Memory); Stops = final-turn Phase-2 stops (Rate = Stops / Items); Near-Orth. = near-orthogonality (∣cos⁡θ∣≤0.1|\cos\theta|\le0.1), measured over adjacent segment pairs (2-Seg) or consecutive segment triples (3-Seg). Segments are sentence-level chunks (sent80), not individual tokens or tool turns.

\text{Model \\& Task}\rlap{K_{\mathrm{osc}}\\;\text{Stop Counts}}\rlap{\text{Near-Orth. (\\%)}}
ItemsStopsRate (%)2-Seg3-Seg
Qwen BFCL v4 Live2,2511,05346.782.710.05
Qwen Agent Memory4657916.991.190.00
Qwen GSM8K1,3191,03178.171.600.01
Gemma GSM8K1,31932024.260.000.00
Qwen FRAMES82477594.051.150.005
  1. Post-hoc stop analysis: Analyzing the stop points with gpt6-astra confirms that 82.19% (263/320) of Gemma GSM8K and 78.48% (62/79) of Qwen Agent Memory stops had already reached an answer or stalled; right/wrong oscillation is 0/320, 1/79 (1.27%), and 0/57 across Gemma GSM8K, Qwen Memory, and Qwen GSM8K disagreement items.
  2. Post-cut completion: Terminating only <think> preserves content to finish arithmetic or correct draft slips (e.g., Gemma GSM8K Problem 182: 30 revised to 23) while keeping total tokens lowest on 4/5 Gemma benchmarks (Comment 2/3).
  3. Multi-turn savings: Across 465 scored Qwen Agent Memory QA sessions (-78.6% tokens over 3×1923\times 192 dialogues), 140 exit in Phase 1, 79 via final-turn KoscK_{\mathrm{osc}} (219 stopped final turns), 245 end naturally, and 1 hits the cap, plus 76 mid-turn KoscK_{\mathrm{osc}} cuts (155 total).
  4. Same-prefix cut vs. continue (CSQA): Stopping beats continuation by +122 Qwen items (85.9% vs. 75.5%) and +248 Gemma items (66.6% vs. 39.8%).
AuthorsOpenReview ↗

Response to Reviewer 6pUU (2/3): Length-Matched Fixed/Random-Late Controls & Full R/C Token Split [W1, Q1]

2/3 — Does trajectory geometry add value beyond a token budget? [W1, Q1]

To test whether shorter <think> alone explains GLOBE's results, we compare length-matched cutoffs on easy GSM8K, medium BFCL v4 Live, and challenging long-horizon Agent Memory, then audit reasoning (RR) and visible-content (CC) tokens. GLOBE reaches 78.2% mean accuracy at 1,065 tokens and 3,590s, outperforming fixed (74.1%) and random (73.0%) Phase-2 cutoffs.

1. Controlled Phase-2 comparison (Qwen3.6-35B-A3B): All policies share Phase 1 (Kt<4K_t < 4, τconf=0.90\tau_{\mathrm{conf}} = 0.90). Conf. Only disables Phase-2 stopping after Kt=4K_t = 4 (no further probing). Fixed-late cuts Phase-2 <think> at constant budget BB (GLOBE's dev late-stop median); Random-late samples cutoffs from GLOBE's target-task late-stop pool (1.01×1.01\times geometric mean on the <think> cutoff scale; realized R+CR+C means 1,118 and 1,121 vs. 1,065). For Agent Memory, Acc scores the 3×155=4653\times 155 = 465 QA sessions while Tok/E2E cover all 3×1923\times 192 dialogues (including 37 memory-writing prerequisites per backend).

Method\bf MethodGSM8K\rlap{\bf GSM8K}BFCL v4 Live\rlap{\textbf{BFCL v4 Live}}Agent Memory\rlap{\textbf{Agent Memory}}Avg.\rlap{\bf Avg.}
Acc ⁠Tok ⁠E2E⁠ ⁠Acc ⁠Tok ⁠E2E⁠ ⁠Acc ⁠Tok ⁠E2E⁠ ⁠Acc ⁠Tok ⁠E2E
GLOBE (Ours)96.57841,96980.03713,07158.02,0405,73078.21,0653,590
Conf. Only96.11,5292,24980.07573,54446.99,18113,03274.33,8226,275
Fixed-late95.87531,95679.63574,43947.12,2456,65474.11,1184,350
Random-late96.17351,92878.83724,75944.12,2556,10673.01,1214,264
  • Single-turn vs. challenging multi-turn: Once Phase 1 screens easy items and controls borrow GLOBE's target-task stop lengths, single-turn gaps are modest on GSM8K (+0.7 vs. Fixed-late, CI 0.00 to +1.52; +0.4 vs. Random-late) and BFCL (+0.4 / +1.2 pts; 90.5% vs. 88.5% / 89.5% on 200-item Non-live; BFCL's 4,439s / 4,759s reflects 131k KV reservation). On challenging multi-turn Agent Memory, blind cutoffs sever active retrieval-synthesis chains, dropping accuracy to 47.1% (-10.9 pts) and 44.1% (-13.9 pts).
  • Ablations & unknown real-world budgets: Disabling Phase 2 (Conf. Only) inflates tokens 3.59×3.59\times (9,181 on Agent Memory, -11.1 pts), and removing τQ\tau_Q lowers accuracy on all 6/6 Paper Table-2 pairs (-0.3 to -5.1 pts). Crucially, both length controls require knowing target-task stop budgets beforehand (~371 on BFCL v4 Live vs. ~2,040 on Agent Memory, a 5.5×5.5\times gap), yet per-problem budgets are unknown beforehand in real deployments; GLOBE requires no prespecified task-specific stop length.

2. Unified 5-benchmark reasoning (RR) and visible-content (CC) token audit: Because GLOBE terminates only <think>, the model can still complete final arithmetic or revise minor draft slips in visible content—raising the question of whether unfinished reasoning merely shifts from RR into CC without reducing total output (Tot\mathrm{Tot}). Below we report <think> (RR), visible response (CC), and Tot\mathrm{Tot} across all five Gemma-4-12B-it Main-Table benchmarks (Agent Memory spans 3×1923\times 192 dialogues across KV, Vector, and RecSum):

Method\bf MethodBFCL v4 Live\rlap{\textbf{BFCL v4 Live}}GSM8K\rlap{\bf GSM8K}CSQA\rlap{\bf CSQA}Character-LLM\rlap{\textbf{Character-LLM}}Agent Memory\rlap{\textbf{Agent Memory}}
R ⁠C ⁠Tot⁠ ⁠R ⁠C ⁠Tot⁠ ⁠R ⁠C ⁠Tot⁠ ⁠R ⁠C ⁠Tot⁠ ⁠R ⁠C ⁠Tot
Full CoT561406011,2301521,3821,6153241,939633747071,3318072,138
DEER447795269145001,414387356743578706481,2107912,001
ThinkBrake310944041,2641541,4191,5053211,826403975001,3167752,091
RCPD371954665694971,0669053631,2689841441,1281,2927862,078
GLOBE (Ours)22877306451267717349340689563816448727531,625

Even where single-turn CC rises slightly (77 vs. 40, 267 vs. 152), GLOBE achieves both the lowest RR and the lowest Tot\mathrm{Tot} on 4/5 benchmarks—BFCL, GSM8K, CSQA (49.1%–64.5% cuts), and Agent Memory (872 / 753 / 1,625, lowest RR, CC, and Tot\mathrm{Tot} across all methods)—and ranks second on Character-LLM (644 vs. 707). On Qwen, GLOBE similarly uses 1,239.4 R+797.3 C=2,036.71{,}239.4\,R + 797.3\,C = 2{,}036.7 on Agent Memory (vs. ThinkBrake 6,958.1+648.9=7,607.06{,}958.1 + 648.9 = 7{,}607.0) and 439.6 R+344.2 C=783.8439.6\,R + 344.2\,C = 783.8 (784 Table-1 total) on GSM8K vs. Full CoT 1,641.3 R+118.8 C=1,760.11{,}641.3\,R + 118.8\,C = 1{,}760.1 (1,761 API total, -55.5%).

AuthorsOpenReview ↗

Response to Reviewer 6pUU (3/3): Transfer Beyond Threshold Selection, R1-Distill-14B & Paired 95% CIs [W2, Q2]

3/3 — Does quality hold beyond the threshold-selection dataset? [W2, Q2]

Below, we distinguish (a) the 9 other model–task pairs in Paper Table 1 from (b) the new model and 3 harder benchmarks evaluated in this rebuttal, and report item-paired 95% CIs. We summarize the key findings and numerical results in this response. For readers interested in the complete experimental tables and additional details, we kindly refer them to our responses to Reviewer nBvC: Comment 3/3 (AIME 2026, LiveCodeBench v6, and FRAMES), Comment 1/3 (DeepSeek-R1-Distill-Qwen-14B), and Comment 2/3 (CoDE-Stop and Conformal Thinking).

Direct Answer: Yes. With Qwen BFCL v4 Live as the sole threshold-selection set, the frozen thresholds (0.90,4,8,0.14)(0.90, 4, 8, 0.14) transfer without retuning to the 9 other Main-Table model–task pairs (-48.4% tokens, -52.5% E2E, +0.74 mean pts, 6/9 matching or exceeding Full CoT), DeepSeek-R1-Distill-Qwen-14B, and three additional hard benchmarks (AIME 2026, LiveCodeBench v6, FRAMES). We also agree that the +0.54 pts in Table 1 should be framed as quality preservation rather than an accuracy gain.

1. Development boundary and zero-retuning transfer:

  • 9 other Paper Table-1 pairs (outside threshold selection): (τconf=0.90,Kphase=4,Kosc=8,τQ=0.14)(\tau_{\mathrm{conf}}=0.90, K_{\mathrm{phase}}=4, K_{\mathrm{osc}}=8, \tau_Q=0.14) was selected on Qwen BFCL v4 Live (N=2,251N=2{,}251; flat 78.63%–79.96% across τQ∈[0.06,0.30]\tau_Q \in [0.06, 0.30], Appendix Table 10), which we will explicitly mark as the threshold-selection set in Table 1 and §4.5 (in controlled factor ablations such as Comment 2/3 and Reviewer E2nw 2/3, it serves only as a fixed reference alongside non-selection benchmarks, while all generalization claims here exclude it). Across the 9 remaining Table-1 pairs (including long-horizon Agent Memory: 3×155=4653\times 155 = 465 scored QA sessions, 3×1923\times 192 total dialogues), frozen thresholds cut tokens by 48.4% and E2E by 52.5%, with 6/9 matching or exceeding Full CoT and the other 3 within 0.2–2.1 pts (+0.74 mean pts).
  • Additional rebuttal evaluations beyond Paper Table 1 (frozen thresholds):
    • Three hard benchmarks (AIME 2026, LiveCodeBench v6, FRAMES): GLOBE alone keeps all 6/6 model–task gaps within 5 pp of Full CoT (75.56% Qwen and 62.98% Gemma means at 1.55×1.55\times and 1.63×1.63\times speedups; Reviewer nBvC 3/3).
    • DeepSeek-R1-Distill-Qwen-14B (Reviewer nBvC 1/3): To cover the DeepSeek-R1-Distill model family commonly evaluated in prior early-stopping papers (DEER, ThinkBrake, RCPD, CoDE-Stop, and Conformal Thinking), we evaluate DeepSeek-R1-Distill-Qwen-14B on all three non-tool benchmarks (CSQA, Character-LLM, GSM8K; its official chat template lacks native multi-turn function-calling tool-schema support in our vLLM runner for BFCL and Agent Memory), where GLOBE ranks first across all 3-task means (76.34% vs. Full CoT 76.30%, 544 tokens [-18.7%], 1,169s [-26.0%]).
  • Comparison with calibrated baselines (Reviewer nBvC 2/3): On Gemma, Conformal Thinking (Wang et al., 2026) requires 400–512 labeled target items per task (4.05×4.05\times slower online, 7.44×7.44\times including calibration), while CoDE-Stop (Hosseini et al., 2026) runs 1.47×1.47\times slower and scores 0.77 mean pts lower.

2. Item-paired 95% CIs outside the threshold-selection set: We frame +0.54 strictly as quality preservation. Diff = GLOBE minus Full CoT (pts); CI Low/High = paired 95% bootstrap bounds (155-scenario cluster bootstrap for Agent Memory); Speedup = Full E2E / GLOBE E2E; NAS = Character-LLM's rubric-based Normalized Acting Score. Paired 95% CIs are strictly positive on Gemma BFCL v4 Live and Qwen Agent Memory, and cross zero with no statistically significant difference detected on the other three benchmarks (direct overthinking evidence is tested via same-prefix cuts in Comment 1/3):

Model and Task\bf Model\ and\ TaskSample and Accuracy\rlap{\textbf{Sample and Accuracy}}Paired CI (pts)\rlap{\textbf{Paired CI (pts)}}Efficiency vs. Full CoT\rlap{\textbf{Efficiency vs. Full CoT}}
ItemsFull (%)GLOBE (%)⁠ ⁠DiffCI LowCI High⁠ ⁠Tok Cut (%)E2E Cut (%)Speedup
Gemma BFCL v4 Live2,25177.5781.12+3.55+2.09+5.0249.152.52.11
Qwen Agent Memory46554.1959.78+5.59+0.43+10.7578.664.92.85
Qwen FRAMES82462.2662.50+0.24-1.70+2.1849.936.51.58
Qwen Character-LLM (NAS)85794.3594.48+0.14-0.91+1.1531.830.71.44
Qwen GSM8K1,31996.4496.51+0.08-0.68+0.8355.543.31.76

Thank you once again for your thoughtful feedback on our paper. If our responses and additional experiments have helped address your concerns, we would sincerely appreciate it if you could consider raising your rating. Please feel free to share any further questions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We thank all reviewers (6pUU, E2nw, nBvC). GLOBE unifies early trial-answer exits and later spherical-trajectory stopping via a training-free phase switch, replacing trial probes with async segment embeddings without target calibration or fixed budgets. We also make live-streaming vLLM + SSE E2E latency a Main-Table metric (the first such systematic early-stopping evaluation at submission time) and evaluate long-horizon tool/memory agents. Detailed tables: reviewer replies (1/3–3/3).

Concern (Reviewers)New experimentKey evidence
Does geometry beat simply cutting reasoning shorter? (6pUU)Keep Phase 1 fixed; compare confidence-only, fixed/random late cutoffs; split reasoning vs. answer tokens (6pUU 2/3).Across GSM8K, BFCL, and Agent Memory: 78.2% mean accuracy, vs. 74.3% / 74.1% / 73.0%; +10.9/+13.9 pp over fixed/random cuts on Memory. GLOBE uses 1,065 tokens, 3,590s E2E (vs. fixed/random 4,350/4,264s) and has the fewest reasoning and total tokens on 4/5 Gemma tasks.
Does it stop productive exploration? (6pUU, E2nw)Test orthogonal reasoning paths, remove the displacement safeguard, and inspect 502 stopped traces across five model–task pairs (6pUU 1/3; E2nw 1/3).The displacement check lets 74.8%–86.8% of still-progressing BFCL/Memory trajectories continue; 78.5%–82.2% of inspected stops had reached an answer or stalled. Only 1/502 showed right/wrong switching; 10-segment orthogonal sequences: 0.00%.
Do frozen thresholds transfer to unseen tasks? (6pUU, nBvC)Select thresholds only on Qwen BFCL v4 Live; apply them without retuning to the 9 other Main-Table pairs and all added benchmarks/models (6pUU 3/3).Across 9 other pairs, 48.4% fewer tokens and 52.5% less E2E, with +0.74 pts mean score (6/9 ≥ Full CoT). Excluding both BFCL pairs as a stricter check: +0.40 pts across 8 cells (5/8 ≥ Full CoT; max drop 2.1 pts). CIs positive for Gemma BFCL and Qwen Memory; otherwise cross zero.
Does it work on harder reasoning tasks? (E2nw, nBvC, 6pUU)Add AIME 2026 (30 problems), LiveCodeBench v6 (175 newest problems), and FRAMES (824); compare early-exit baselines on Qwen/Gemma (nBvC 3/3; E2nw 3/3).GLOBE leads early-exit baselines in mean accuracy: 75.56% on Qwen (1.55× E2E speedup) and 62.98% on Gemma (1.63×); all 6/6 model–task results are within 4.57 pp of Full CoT.
How does real serving compare with recent baselines? (nBvC, 6pUU)Compare CoDE-Stop and Conformal on five Gemma tasks; include online and calibration time (nBvC 1/3–2/3).GLOBE has lowest E2E on all 5/5 tasks (3,001s, 78.08% mean score). CoDE-Stop is 1.47× slower (77.31%); Conformal is 4.05× slower online, or 7.44× including calibration (78.58%).
Does it transfer to another reasoning model? (nBvC, 6pUU)Test DeepSeek-R1-Distill-Qwen-14B on CSQA, Character-LLM, GSM8K without retuning (nBvC 1/3).GLOBE ranks first in three-task mean score (76.34% vs. Full CoT 76.30%), using 18.7% fewer tokens and 26.0% less E2E (544 tokens; 1,169s).
Is the embedding overhead practical and robust? (E2nw, nBvC)Compare 8B/4B/0.6B embedding models on four tasks; test other embedding families and latency (E2nw 2/3; nBvC 3/3).Even 0.6B (1.2 GB) achieves 80.24% vs. 81.80% with the 8B embedding model, +8.43 pp over RCPD and 2.0–2.2× speedup vs. Full CoT/DEER/ThinkBrake; 8.15–27.87 ms median embedding latency (80–320 tok) and 89.3%–98.8% stop/continue agreement across BGE, E5, and MPNet.
Why segment reasoning instead of using long contexts? (nBvC)Compare sentence segments (sent80) with expanding prefixes, sliding 512-char, and random windows (nBvC 3/3).Expanding/sliding windows mask local reversals (0 stops); sent80 catches post-solution re-checking (11 stops; 9/11 correct; 8/11 judged ready to stop, vs. 0/1 for random windows).

Individual replies detail J≥10J\ge10, τQ\tau_Q ablations, Qwen Rescue/Harm (GSM8K +14/-12; Memory +51/-28), raw counts, and §4.3 wording.

Planned revisions (and code release): §4.1–4.2 Main Results: add CoDE-Stop/Conformal, AIME 2026 / LiveCodeBench v6 (175 newest problems; broader coverage in revision) / FRAMES, and R1-Distill-14B; identify Qwen BFCL v4 Live as the sole threshold-selection set (§4.5) and report zero-retuning transfer with CIs. §3.3, 4.3: add geometry definitions, length controls, reasoning/content token splits, and segmentation ablations. §4.4 and Appendices B/D: add trajectory audits, counts, embedding-model robustness, and latency benchmarks.

We thank the reviewers again. Before November 18 (AoE), we welcome questions and would appreciate knowing whether these results address your concerns.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.