acceptodds
Withdrawn from ICLR 2027

ZS-MTR-RR: Training-Free Length-Conditioned Reliability Routing of Global and Multi-Trajectory Evidence for Zero-Shot AI-Generated Text Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Large language models increasingly produce fluent, human-like text, posing severe challenges to academic integrity, misinformation governance, and content provenance. Zero-shot detectors avoid collecting labeled data for unseen generators, yet existing approaches face a dilemma: document-level statistics compress token probabilities into a single global score, overlooking positional sequence structure; while trajectory-based methods retain positional information but rely on fixed aggregation, failing to adapt the relative contributions of global and trajectory evidence as text length varies. To address these limitations, we propose ZS-MTR-RR, a training-free framework featuring Zero-Shot Multi-Trajectory Representation with Length-Conditioned Reliability Routing. Using paired frozen base and instruction-tuned models, ZS-MTR-RR extracts aligned token log-probability outputs to construct both a full-document proxy-discrepancy path and a multi-trajectory evidence path capturing late-stage stability, proxy tension, and token-reuse structures. Conditioned on scored length as an indicator of available positional support, a deterministic soft gate balances the two paths, converting fixed signal combination into length-conditioned reliability routing without training an additional detector or learned router. We evaluate ZS-MTR-RR on the five-domain EvoBench benchmark and four RealBench-derived transformations. Under the unified EvoBench protocol, ZS-MTR-RR achieves the highest cross-domain average AUROC (reaching 0.8753) among evaluated methods. Ablation and length-stratified analyses show that the two evidence paths provide complementary discriminative signals and support the benefit of length-conditioned routing over fixed fusion on the primary benchmark.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI-generated text detection, zero-shot detection, multi-trajectory representation, reliability routing, length-conditioned routing, training-free

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_drCEOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes ZS-MTR-RR, a zero-shot AI-generated text detection method that adaptively combines global text features with trajectory features based on the input text length. The proposed method also uses a model pair, one base model and one instruction-tuned model, to strengthen the feature set through comparative analysis. Evaluation on existing benchmarks demonstrates that the proposed method outperforms several existing baselines.

Critical strengths
  1. This paper proposes an adaptive fusion method that smoothly routes the score from the global score to the trajectory score based on text length, making detection performance more stable across different text lengths.
  2. Evaluation on existing datasets demonstrates the better performance of the proposed method against existing methods.
Critical weaknesses
  1. The novelty of this paper needs more clarification. Several important components and insights have already been well explored by existing studies. For example, the base and instruction-tuned model pair is directly adopted from Binoculars [1]; the token-reuse features HH and GG, which evaluate the output distribution at a position against the token just read, are essentially the backward loss of BiScope [2]; the multi-lag recurrence EE is very similar to the context window in Profiler [3]; and the surprisal-volatility feature VV together with its information-density interpretation originates from GPT-who [4]. I highly suggest the authors investigate related work more thoroughly and re-position the paper correctly within the existing literature.
  2. Several key designs of the paper need more justification; they appear empirical or arbitrary rather than principled. For example, why does HH use the top-100 candidates rather than a ratio of the vocabulary size? Why does GG use a uniform estimate for tokens outside the top-100? The method only ever uses the probability and rank of the token just read, so the authors could store these two numbers directly instead of caching all top-100 probabilities and approximating the tail. Additionally, why does EE use token recurrence rather than word recurrence? No ablation or theory justifies these choices.
  3. The global score DzD_z performs much worse than the Binoculars or Fast-DetectGPT score, although all of them are computed from the same proxy outputs. Why not replace DzD_z with one of these better-performing global scores? This would not affect the paper's claimed advantage of adding the trajectory score to the final score.
  4. Many key hyper-parameters of the proposed method are selected empirically and need calibration for different domains or models, so out-of-distribution evaluation is essential for testing the generalizability of the proposed method when deployed. Appendix B.2 further admits that the feature set, coefficients and gate were developed with reference to EvoBench results and that the scalers are estimated on the evaluation pool, so the main results are optimistically biased relative to the baselines. Thus, instead of only showing AUROC, the authors should report the F1 score of the proposed method against baselines with a fixed threshold.
  5. According to Table S11, the proposed method shows more vulnerability to adversarial attacks than several baselines, which may expose more problems when deployed.

[1] Hans, Abhimanyu, et al. "Spotting llms with binoculars: Zero-shot detection of machine-generated text." ICML. 2024.

[2] Guo, Hanxi, et al. "Biscope: Ai-generated text detection by checking memorization of preceding tokens." NeurIPS. 2024.

[3] Guo, Hanxi, et al. "Profiler: Black-box AI-generated Text Origin Detection via Context-aware Inference Pattern Analysis." EMNLP. 2025.

[4] Venkatraman, Saranya, Adaku Uchendu, and Dongwon Lee. "Gpt-who: An information density-based machine-generated text detector." Findings of NAACL. 2024.

Questions for authors

Addition to the weaknesses, here are more detailed questions:

  1. In the complementary evidence, both DD and AA are computed from the difference of the two log-probability sequences (DD on the full sequence, AA on the latter half). Why is there no corresponding counterpart for VV?
  2. Why does the proposed method only consider the latter half, rather than combining it with the former half or other finer-grained ranges? The experiments show that the latter half performs best on its own, but other ranges may be complementary.
Reviewer_JDJYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper claims that zero-shot AI-generated content detection should consider the context length and adjust the contribution of different detection signals according to the length. Based on this idea, the work proposes a routing mechanism to fuse global signals with multiple trajectory features. The routing is implemented as a length-conditioned gate that requires no additional training, together with an evaluation of whether adjusting the weights by length improves over fixed-weight fusion. This work did a comprehensive evaluation for the proposed method and shows good performance on various settings.

Critical strengths
  1. This paper investigates the affect of context length for LLM-generated content detection and explores the detection performance on long and short context with different detection metrics. This work addresses the limitation of the exisiting works that use a fixed combination of signals without adjusting their contributions according to text length. Based on the observation, this paper designs a novel fusion strategy to handle this task.

  2. The authors use a simple gate to vary the fusion weights with text length and obtain performance gains on EvoBench. The experiments compare the method not only with individual paths but also with fixed fusion weights selected on the development set, which makes the evidence for the gate's own contribution more convincing.

  3. The ablation study is comprehensive and clear. The authors conduct a fairly extensive set of component ablations on zero-shot evidence such as stability, late-stage model discrepancy, and repetition features, helping readers understand where the gains of the full combination come from.

Critical weaknesses
  1. As the main contribution of this work, the design of gating module is not clearly stated in the paper. Specifically, how the default values 52 and 6 are originally determined is not explained.

  2. The analysis for context length is not strongly convincing. The authors mainly analyze four scored-length intervals—≤32, 33–52, 53–128, and >128 tokens—with an input limit of 512 tokens. I am concerned that this range still mainly covers relatively short texts, and that the broad >128 group does not adequately show how performance changes on longer texts. Moreover, these are also scored-length intervals after truncation; reporting the original text-length distribution would help clarify what kinds of texts the experiments actually cover.

  3. The related work section does not sufficiently discuss the length-dependence analyses and global/trajectory fusion already explored by the closest prior methods. For example, Sun et al [1], already discuss the limitations of temporal signals on short texts and their complementarity with global signals, and suggest dynamically adjusting fusion weights by length in their Limitations section.

[1] Sun K, Bao G, Cui H, et al. When ai settles down: Late-stage stability as a signature of ai-generated text detection[J]. arXiv preprint arXiv:2601.04833, 2026.

Questions for authors
  1. Why does losing positional structure matter for AI-generated text detection? A document-level mean loses ordering and local variation, but I would like a clearer explanation of why this information can reliably distinguish human-written from AI-generated text.

  2. Why can the likelihood difference between base and instruction-tuned models serve as a signal for AI-generated text detection? To what extent does its effectiveness depend on the choice of instruction-tuned model, or on the data and training procedure used during SFT?

  3. The paper divides scored length into ≤32, 33–52, 53–128, and >128 tokens; how were these intervals chosen? Do they come from previous experimental settings, the distribution of the data, or an empirical choice made in this paper?

  4. The paper retains only the first 512 tokens of longer texts and keeps shorter texts in full. Is this intended to provide a common evaluation budget, or does it follow a common setting in previous detection work?

Reviewer_HyaKOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ZS-MTR-RR, a training-free AI-generated text detector that combines global model discrepancy with handcrafted trajectory features through a length-dependent gate. The method achieves competitive results on EvoBench and is evaluated through ablations, external transformations, and alternative proxy models.

Critical strengths

The paper achieves competitive results on EvoBench and provides extensive experiments, including component ablations, length-stratified analyses, source-disjoint validation, and evaluations on transformed text and alternative proxy models. These analyses help assess the contributions of individual components and clarify the method’s applicability and limitations.

Critical weaknesses
  1. Limited methodological novelty. The core signals, including base–instruction model discrepancy, late-stage volatility, and repetition-probability cues, have been explored in prior work. The final detection score is

    SRR(x)=g(L)Dz(x)+[1−g(L)][Vz(x)+Az(x)+0.2(Hz(x)+Gz(x)+Ez(x))](R1)S_{\mathrm{RR}}(x)=g(L)D_z(x)+[1-g(L)]\bigl[V_z(x)+A_z(x)+0.2\bigl(H_z(x)+G_z(x)+E_z(x)\bigr)\bigr]\qquad\text{(R1)}

    where g(L)=1/[1+exp⁡((L−52)/6)]g(L)=1/[1+\exp((L-52)/6)] depends only on scored text length. The proposed contribution mainly consists of adapting existing signals, adding handcrafted statistics, and combining them through this length-dependent weighting scheme. While this combination provides empirical improvements, the methodological advance beyond existing detection signals appears incremental.

  2. Inconsistent benefits of length-conditioned routing. Length-conditioned routing is presented as a central methodological contribution, yet its benefits are inconsistent across evaluation settings. Although it improves results on EvoBench, it underperforms static fusion on external transformations (Table S12) and degrades performance relative to the trajectory-only baseline with Llama-3.1 proxies (Table S13). These results weaken the evidence for the generalizability of this central contribution.

  3. Insufficient justification of specific heuristic choices. Several choices in feature fusion and hyperparameter selection need stronger justification. For the trajectory score in Eq. (R1), it is unclear why the five features are combined with weights 1,1,0.2,0.2,0.21,1,0.2,0.2,0.2. Median/MAD normalization makes their scales comparable, but does not account for differences in discriminative strength or redundancy. With only five inputs, a linear model or a simple MLP could learn their combination with little to no additional inference cost, especially since labeled development data are already used for gate selection. Additionally, although region ablations and gate-selection experiments are provided, other settings—including the top-100 reuse cutoff, volatility-window radius of 10, and recurrence-lag cap of 32—lack clear selection criteria and sensitivity analysis. These hyperparameters could also be selected through cross-validation on source development data.

  4. The paper is not well written or carefully proofread. The writing often piles up technical terms and uses long sentences, starting in the abstract (e.g., the sentence beginning “Using paired frozen base and instruction-tuned models,” p. 1, lines 23–26). Several sentences contain multiple claims with references grouped at the end, making it difficult to identify which reference supports which claim (e.g., p. 1, lines 50–53, and p. 3, lines 151–154). Narrative citations are also used where parenthetical citations would be appropriate (e.g., p. 1, lines 39–41 and 45–46, and p. 3, lines 125–126). The writing and citations should be carefully checked throughout the paper.

Questions for authors
  1. The AI Use Statement mentions assistance with “theoretical derivations,” but the manuscript does not include any corresponding theoretical analyses. Could the authors specify which derivations this refers to and what assistance AI provided?
  2. The paper does not compare inference time with other detectors. Given the simplicity of the method, if it is faster than existing approaches, the authors could demonstrate this advantage.
Reviewer_YY3iOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper combines global likelihood discrepancy with late-stage stability, proxy tension, and token-reuse features from frozen base and instruction-tuned models. After median/MAD normalization, a deterministic length gate mixes the global and trajectory scores. On EvoBench, the method reaches 87.53% macro AUROC, compared with 85.19% for Binoculars. Ablations support the combined representation, while retrospective source-disjoint experiments show a smaller gain from routing over selected static fusion. Transformation and proxy experiments also identify limits to this routing benefit.

Critical strengths
  1. The method is straightforward to implement with a frozen model pair. Both paths reuse the same two forward passes, without perturbation generation or a fitted detector. This makes the proposed combination a practical extension of existing statistical detectors.
  2. The EvoBench improvement is supported by more than an aggregate score. Table 1 reports a 2.34-point macro AUROC gain over Binoculars, with a 95% source-cluster bootstrap interval of [1.52, 3.15]. The source-disjoint comparisons also favor the full representation over V+A routing and selected routing over static fusion (Table S3). These controls support a useful empirical contribution.
Critical weaknesses
  1. The complete design lacks an evaluation independent of its development. Appendix B.2 states that feature choices, coefficients, and the default gate were developed using EvoBench results. The retrospective folds isolate sources during fusion selection, but retain those earlier design choices. The transformation panel was also selected during development. The authors disclose this clearly; nevertheless, the reported advantage may include selection effects that the bootstrap intervals do not capture.
  2. The central routing benefit does not transfer consistently. With fixed source scales, RR obtains 81.69% AUROC on the transformation panel versus 81.89% for static fusion, and is essentially tied with trajectory-only scoring (Table S12). Routing also reduces AUROC relative to trajectory-only scoring by 1.46 points for Llama-3.1-8B (Table S13). Since most pooled gains over trajectory-only scoring on EvoBench arise from cross-length comparisons (Appendix D.1), it remains unclear when the gate improves evidence selection beyond adjusting score comparability across lengths. This limits the case for routing as the main contribution.
Questions for authors
  1. Independent validation: Is there an evaluation set that did not inform feature, coefficient, gate, or benchmark selection? If available, results with the complete configuration and source scalers frozen, compared with static fusion and trajectory-only scoring, would directly address my main concern.
  2. Routing mechanism: Do the global and trajectory paths retain the proposed short-text/long-text crossover on the transformation panel and with Llama-3.1? A comparison against static fusion with length-dependent score normalization, using only development data to set that normalization, would help distinguish routing gains from score-scale effects.

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

Related papers

Discussion (0)

Sign in to comment.