PaperlayerView exact revision on arXiv

Source-linked AI summary

Antislop: A Comprehensive Framework for Identifying and Eliminating Repetitive Patterns in Language Models

Samuel Paech, Allen Roush, Judah Goldfeder, Ravid Shwartz-Ziv

arXiv:2510.15061v2cs.LGcs.CL

TL;DR

LLM outputs contain repetitive patterns that can appear far more often than in human writing, motivating tools that detect and suppress model-specific slop. Antislop combines backtracking-based sampling, automated preference-data generation, and FTPO. Across the reported tests, the sampler and FTPO achieved strong suppression with negligible or no measurable quality loss, while the sampler’s throughput cost remains a limitation.

  • Problem

    Some repetitive LLM patterns occur over 1000× more frequently than in human writing, while token banning can cause collateral damage.

    Source: §1 · p. 1

  • Method

    Antislop combines backtracking-based sampling with an automated pipeline and FTPO, which makes targeted token-level adjustments using generated preference data.

    Source: §1 · p. 2, §7 · p. 10

  • Results

    FTPO outperforms DPO on the measured metrics, while the Antislop Sampler achieves 100% suppression without quality loss.

    Source: §6.2 · p. 7, §7 · p. 10

  • Takeaways & Limitations

    The framework supports model-specific slop suppression while preserving evaluated output quality and capabilities.

    Source: §7 · p. 10, §8 · p. 10

    Frequent sampler backtracking decreases performance by 69%-96% depending on banlist size.

    Source: §4.3 · p. 5

Abstract

from arXiv · show

Widespread LLM adoption has introduced characteristic repetitive phraseology, termed "slop," which degrades output quality and makes AI-generated text immediately recognizable. We present Antislop, a comprehensive framework providing tools to both detect and eliminate these overused patterns. Our approach combines three innovations: (1) The Antislop Sampler, which uses backtracking to suppress unwanted strings at inference time without destroying vocabulary; (2) An automated pipeline that profiles model-specific slop against human baselines and generates training data; (3) Final Token Preference Optimization (FTPO), a novel fine-tuning method that operates on individual tokens, surgically adjusting logits wherever a banned pattern has appeared in an inference trace. We demonstrate that some slop patterns appear over 1,000x more frequently in LLM output than human text. The Antislop Sampler successfully suppresses 8,000+ patterns while maintaining quality, whereas token banning becomes unusable at just 2,000. Most importantly, FTPO achieves 90% slop reduction while maintaining or improving performance in cross-domain evals including GSM8K, MMLU, and creative writing tasks. In contrast, DPO suffers significant degradation in writing quality and lexical diversity despite achieving weaker suppression. We release all code and results under MIT license: https://github.com/sam-paech/auto-antislop.

1 Introduction

Antislop targets repetitive “slop” in language-model outputs, which can occur over 1000× more often than in human writing. It combines inference-time backtracking with FTPO to suppress these patterns while limiting quality damage.

  • Overused phrases can occur over 1000× more frequently in LLM text than human writing, making repetition recognizable as AI-generated “slop.”Source: §1 · p. 1
  • Existing token banning can create collateral damage because banning a tokenized prefix also bans unrelated words sharing that prefix.Source: §1 · pp. 1–2
  • The Antislop Sampler detects unwanted words, phrases, and regex patterns, backtracks to their first token, lowers its probability, and resamples.Source: §1 · p. 2
  • The sampler can suppress 8,000 patterns with configurable strength, from soft discouragement to hard banning, without degrading output.Source: §1 · p. 2
  • FTPO is designed to suppress slop surgically while minimizing collateral changes to the model.Its motivation is that large logit adjustments can damage the model when disfavoring highly preferred tokens.Source: §1 · p. 2

2 Related Work

Prior work addresses repetition through decoding constraints, lexical filtering, unlikelihood training, and preference optimization, but these approaches target different failure modes. Antislop is positioned against limitations involving context-specific repetition, statistical overuse, diversity, and collateral effects.

  • Stochastic decoding increases diversity and reduces incoherence, but does not address repetitive tendencies in otherwise coherent outputs.Source: §2 · p. 2
  • XTC removes high-probability tokens and DRY blocks repeated sequences, but neither directly identifies statistically overused patterns across independent generations.Source: §2 · p. 2
  • Beam-search constraints and lexical filters enforce forbidden outputs, while related methods examine backtracking or omit selected characters without fine-tuning.Source: §2 · p. 2
  • Unlikelihood training penalizes unwanted continuations, including repetitive loops and disfavored tokens or n-grams.Source: §2 · p. 2
  • DPO aligns models on preference pairs but has reported failure modes including reduced preferred-response likelihood, diversity collapse, and lower syntactic and n-gram variety.Source: §2 · p. 3

3 Forensic Analysis of Over-represented Patterns

The forensic analysis identifies model-specific slop by comparing pattern frequencies in generated creative writing with human baselines. It finds extreme overrepresentation of names, clichés, descriptive terms, and sentence constructions, with fingerprints varying across model families.

3.1 Quantifying Slop

  • 3.1 Quantifying Slop: The analysis generates 2,000 creative-writing outputs per model and computes word, bigram, and trigram frequency ratios against human corpora.Source: §3.1 · p. 3
  • 3.1 Quantifying Slop: Collated overrepresented words and n-grams form a model-specific “slop fingerprint.”Source: §3.1 · p. 3

3.2 Empirical Findings

  • 3.2 Empirical Findings: “Elara” appears 85,513 times more often in gemma-3-12b outputs than human text, while “heart hammered ribs” is 1,192× overrepresented.Source: §3.2 · p. 3, Table 1 · p. 3
  • 3.2 Empirical Findings: Slop fingerprints cluster within model families but differ between families, motivating model-specific identification and suppression.Source: §3.2 · p. 3
  • 3.2 Empirical Findings: Some models use “It’s not X, it’s Y” 6.3× more often than human writing, alongside recurring names, sensory clichés, intensifiers, and descriptives.Source: §3.2 · p. 3

4 The Antislop Sampler

The Antislop Sampler detects complete unwanted sequences, backtracks to their start, and resamples with configurable suppression strength. This preserves alternatives and coherence, but repeated backtracking can impose substantial throughput costs.

4 The Antislop Sampler

  • 4 The Antislop Sampler: The sampler detects words, phrases, and regex patterns only after the complete sequence appears, reducing false positives relative to token banning.Source: §4 · p. 4

4.1 Backtracking Mechanism

  • 4.1 Backtracking Mechanism: It scans the token trace, backtracks to the pattern start, reduces the initiating token’s probability, and resamples using min-p filtering.Source: §4.1 · p. 4, Figure 2 · p. 4
  • 4.1 Backtracking Mechanism: If the same token reappears, subsequent violations are ignored to prevent infinite loops, enabling soft-banning.Source: §4.1 · p. 4

4.2 Soft Banning: Configurable Suppression Strength

  • 4.2 Soft Banning: Ban strength s ranges from free allowance at s = 0 to complete blocking at s = 1, with intermediate values providing incremental suppression.Source: §4.2 · pp. 4–5
  • At ban-strength < 1.0, banned patterns can still pass when their probability is high enough compared to the next highest token, allowing “tapestry” when directly requested in the prompt.Source: §4.2 · p. 5

4.3 Implementation and Limitations

  • 4.3 Implementation and Limitations: Frequent backtracking decreases performance by 69%-96% depending on banlist size, motivating FTPO for speed-sensitive applications.Source: §4.3 · p. 5

5 Final Token Preference Optimization (FTPO)

FTPO suppresses unwanted patterns by training on token-level preference pairs while constraining changes to preserve model behavior. Its formulation combines margin-based preference learning, targeted and non-target regularization, and automated data generation from Antislop Sampler backtracking.

5.1 Limitations of Direct Preference Optimization

  • FTPO addresses DPO’s update-control trade-off by constraining logits to reference values and avoiding large divergences that can degrade or collapse the model.DPO’s β parameter can impair learning at high values or permit damaging logit shifts at small values.Source: §5 · p. 5, §5.1 · p. 5

5 Final Token Preference Optimization (FTPO)

  • FTPO trains on a single continuation token at the end of an inference trace, pairing the rejected first token of a banned sequence with multiple coherent alternatives.The pair includes the prompt and response prefix, one rejected token, and a set of chosen continuation tokens.Source: §5 · p. 5

5.2 The FTPO Formulation

  • FTPO combines a margin-based preference loss with separate regularization for target and non-target logits.Target logits may move within a zero-penalty window, while the remaining vocabulary is anchored to prevent distribution drift.Source: §5.2 · p. 6

5.3 Key Design Principles

  • Logit-space MSE localizes updates to chosen and rejected tokens, while margin deactivation and two-part regularization limit overtraining and destructive divergence.The preference weight becomes zero once the chosen tokens exceed the rejected token by margin m.Source: §5.3 · p. 6

5.4 Automated Training Data Generation

  • The Antislop Sampler automatically generates FTPO preference data by capturing rejected tokens and viable alternatives whenever backtracking detects a banned sequence.This creates an end-to-end pipeline from pattern identification through training-data generation and FTPO training.Source: §5.4 · p. 6

6 Experimental Evaluation

Across six evaluation subsections, Antislop is tested for suppression, quality preservation, capability retention, overtraining robustness, regex control, and hyperparameter sensitivity. FTPO and the Antislop Sampler generally suppress unwanted patterns while preserving quality better than DPO and token banning.

6.1 Experimental Setup

  • The evaluation covers Gemma-3-12B, Mistral-Small-3.2, and Llama-3.3-70B using suppression, writing-quality, diversity, MMLU, and GSM8K measures across 1k–8k-pattern banlists.The primary FTPO and DPO training experiments use Gemma-3-12B, with one epoch and early stopping at 0.85 preference accuracy.Source: §6.1 · p. 7, Table 2 · p. 9

6.2 Main Results: Suppression Performance vs. Writing Quality

  • The Antislop Sampler achieves perfect suppression while improving writing quality, whereas token banning collapses quality to 28 at 8k patterns.DPO drops 6–15 writing-quality points while achieving only 80–82% suppression.Source: §6.2 · p. 7
  • 90% suppression with less than 1% quality loss is achieved by FTPO, while DPO reaches 80% suppression with 15% quality degradation.The comparison is reported for the main suppression-versus-writing-quality evaluation.Source: §6.2 · pp. 7–8

6.3 FTPO vs DPO: Detailed Comparison

  • FTPO suppresses 90+% of slop for banlists up to 4,000 items with negligible effects on writing quality, lexical diversity, GSM8K, and MMLU.It achieves 8.5% stronger suppression than DPO and preserves GSM8K and MMLU within 1–3% of baseline, while DPO degrades both by 2–5%.Source: §6.3 · p. 8, Table 2 · p. 9
  • FTPO maintains baseline-level long-form writing and lexical diversity, while DPO degrades extended creative writing and reduces diversity to 74–92% of baseline.FTPO’s diversity remains at 95–102% of baseline across the compared settings.Source: §6.3 · pp. 8–9

6.4 Robustness to Overtraining

  • FTPO reaches nearly 100% preference accuracy with minimal degradation, while DPO reaches only 40% before substantial degradation occurs.Increasing DPO β to 1.0 mitigates degradation but reduces ban suppression by 15.9%.Source: §6.4 · p. 9, Figure 4 · p. 10

6.5 Regex Bans

  • The Antislop Sampler suppresses 100% of the variable sentence-level pattern “It’s not X, it’s Y” using regex bans.Source: §6.5 · p. 9

6.6 FTPO Hyperparameter Ablations

  • FTPO hyperparameter safeguards are supported by ablations in which sub-optimal settings produce poor preference accuracy and degradation.The tested safeguards include a reference-tether strength and a margin that turns off gradients for winning chosen logits.Source: §6.6 · p. 9, §6.4 · p. 9

7 Discussion

Antislop combines inference-time suppression and targeted training to reduce overused stylistic patterns while preserving measured quality. The sampler achieves complete suppression without quality loss, whereas its throughput cost motivates training-based suppression and future evaluation beyond creative writing.

  • 100% suppression of over-used patterns is achieved by the Antislop Sampler without quality loss.Source: §7 · p. 10
  • 69-96% throughput reduction limits the Antislop Sampler in performance-sensitive deployments because frequent backtracking restarts inference.This creates an incentive to prefer suppression trained into the model weights.Source: §7 · p. 10
  • FTPO is designed to make suppression permanent through targeted training with minimal changes to the model’s remaining distribution.Its minimal degradation is attributed to multi-part regularization and gradient nullification when chosen tokens exceed the margin.Source: §7 · p. 10
  • Future work should test Antislop beyond creative writing, replicate quality metrics with human raters, evaluate AI-text detection, and examine toxic-text suppression.Source: §7 · p. 10

8 Conclusion

The paper introduces Antislop as a framework for detecting and suppressing overused stylistic patterns while preserving capabilities on evaluated benchmarks. Its sampler and FTPO training achieve higher suppression than DPO and token banning with negligible measured quality loss.

  • Antislop eliminates overused stylistic patterns while preserving capabilities on the evaluated benchmarks.Source: §8 · p. 10
  • The framework combines sequence-level backtracking, model-specific slop profiling against human baselines, automated preference-data synthesis, and FTPO training.Source: §8 · p. 10
  • FTPO and the sampler achieve higher suppression than DPO and logit-based token banning with negligible measurable quality loss on the authors’ rubric.Source: §8 · p. 10

Reproducibility Statement

The paper provides the materials needed to reproduce Antislop Sampler and FTPO/DPO training experiments. It documents algorithms, configurations, data processing, evaluation prompts, and inference implementation details.

  • Reproduction materials include algorithms, loss definitions, hyperparameters, training configurations, LoRA settings, stopping criteria, and decoding parameters.Source: Reproducibility Statement · p. 11
  • The release includes the data pipeline, prompts, judge rubric, scoring template, inference implementation, supplementary code, and example configuration files.Source: Reproducibility Statement · p. 11

Ethics Statement

The study uses public datasets and benchmarks without collecting human-subject data or personally identifying information. It identifies potential harms from suppressing legitimate styles or enabling AI-text-detection evasion and proposes controls for deployment.

  • The study uses publicly available datasets and benchmarks, processes only public text, collects no personally identifying information, and requires no IRB review.Source: Ethics Statement · p. 11
  • Potential harms include unintended suppression of legitimate dialects or minority styles and attempts to evade AI-text detection.Source: Ethics Statement · p. 11
  • Mitigations include human-vetted banlists, exposed ban-strength controls, soft-ban defaults, whitelists, and human review of production banlists.Source: Ethics Statement · p. 11
  • The authors report throughput impacts for energy-cost accounting and declare no conflicts of interest or biasing external sponsorship.Source: Ethics Statement · p. 11

Appendices

The appendices evaluate soft banning, inference throughput, and long-range regex enforcement, showing that Antislop can preserve contextually necessary banned terms, suppress unwanted patterns, and eliminate a targeted stylistic pattern.

A Soft Banning

  • 90% suppression in normal generation while fully permitting requested patterns demonstrates the Antislop Sampler’s soft-banning balance at s = 0.4.The mechanism allows banned terms when alternatives are inadequate; hard banning corresponds to ban-strength 1.0.Source: A Soft Banning · p. 15

B Inference Performance (tok/s)

  • The sampler incurs substantial throughput penalties because backtracking increases with banlist size, with additional loss in the chunkwise API implementation.Direct vLLM integration is identified as a possible optimization.Source: B Inference Performance (tok/s) · p. 16
  • 69% lower throughput at banlist size 1,000 and 96% lower throughput at size 8,000 quantify the sampler’s worst-case inference cost.These measurements use optimal parallelism with vLLM; the 8,000-pattern setting is described as a stress test.Source: B Inference Performance (tok/s) · p. 16, Figure 6 · p. 16

C Long-range constraint enforcement via regex bans

  • Regex backtracking scans generated text, maps the earliest offending span to a token, and resamples from cached top-logprob lists without another API call.The procedure preserves the decoding hyperparameters while enforcing long-range constraints.Source: C Long-range constraint enforcement… · p. 17
  • 0.00 hits for the “not x, but y” family with Antislop qwen3-4b contrasts with the baseline model’s overuse of the pattern.Figure 7 reports the outcome for regex-based enforcement.Source: Figure 7 · p. 17, C Long-range constraint enforcement… · p. 17

D Hyperparameter Ablations

This section examines FTPO’s hyperparameter choices and ablations, including clipping, reference tethering, loss components, and regularization. It also details the configuration of slop controls and related analyses of suppression, quality tradeoffs, model-specific patterns, and iterative generation.

D Hyperparameter Ablations

  • 16 for clip_epsilon_logits effectively disables protection, causing model collapse into single-word repetitions, whereas the default 2 preserves writing quality at 95% preference accuracy.The collapsed model’s logits diverged much farther from reference.Source: D Hyperparameter Ablations · p. 18
  • 71% writing-quality degradation occurs when lambda_mse_target is 0; 0.4 limits learning to 74% preference accuracy, while the default 0.05 reaches 95% without substantial degradation.Removing the target-logit reference tether speeds training and increases divergence, whereas strengthening it constrains divergence but impairs preference accuracy.Source: D Hyperparameter Ablations · p. 18

E DPO β Hyperparameter Ablation

  • 94.4% suppression with negligible (< 1%) longform-writing degradation is achieved by FTPO, while DPO at β = 1.0 limits degradation below 1% but suppresses only 74.7%.At lower DPO β values, writing quality is markedly reduced.Source: Figure 8 · p. 19

J FTPO Loss Function Definition

  • FTPO combines a margin-based preference loss with MSE tethering terms so models learn token preferences while limiting uncontrolled drift of the logit distribution.The target MSE term includes a zero-penalty window around the reference logits.Source: J FTPO Loss Function Definition · p. 26

K Slop Profile Clustering Between Models

  • Across tested models, “flickered” appears on 98.5% of top over-represented lists, while trigram “voice barely whisper” appears on 68.7% of lists, showing high cross-model correlation in these patterns.The method identifies over-represented usages and is domain-specific; over-used patterns in creative writing may differ from those in professional writing.Source: K Slop Profile Clustering Between Models · p. 27

M Auto-antislop Configuration File for gemma-3-12b-it 2k Banlist Size

  • The auto-antislop pipeline iteratively generates outputs, bans identified strings and n-grams, recomputes slop, and adds newly emerging patterns in later iterations.The configuration specifies a minimum of 2 iterations to catch slop that emerges after the initial banlist is applied.Source: M Auto-antislop Configuration File for… · p. 29
  • The configuration combines dictionary and non-dictionary n-gram quotas, over-represented-word analysis, whole-phrase bans, user-defined bans, whitelists, and regex patterns.It sets initial quotas such as 300 dictionary bigrams, 200 non-dictionary bigrams, 300 dictionary trigrams, 200 non-dictionary trigrams, 920 dictionary over-represented words, and 80 non-dictionary words.Source: M Auto-antislop Configuration File for… · pp. 32–33
  • The fine-tuning setup enables FTPO with a 2,500-token maximum sequence length, 4-bit loading, LoRA rank 256, five unfrozen layers, and a 0.85 early-stopping threshold.Source: M Auto-antislop Configuration File for… · p. 35
  • FTPO uses a 0.8 rejected-sample regularization strength and 0.2 chosen-sample strength to address skew in generated training data.The configuration notes that strong rejected-sample downsampling is useful because raw generated datasets are typically highly skewed.Source: M Auto-antislop Configuration File for… · p. 36
  • FTPO computes separate losses for chosen and rejected target tokens versus the remainder of the vocabulary, allowing more freedom for non-target logits.The target-token MSE tether uses lambda 0.05 with a 0.5-logit grace bandwidth, while the configuration also defines a 2-logit preference-loss clipping margin.Source: M Auto-antislop Configuration File for… · pp. 36–37

Related papers

Loading 2510.15061v2…