文章背景与核心概要
大语言模型(LLM)在面对冗长的长文本事实枚举任务(如列举电视剧集、星座或宝可梦)时,经常会出现一种可复现的失效模式:陷入严重的重复循环(如严密的逐字循环或退化的列表)。这种现象不仅严重影响了生成质量,也阻碍了模型在复杂场景下的可靠应用。
本文深入探讨了是否可以通过定向权重编辑(例如修改单个符号反转的神经元)来解决这些循环问题,而无需依赖传统的重复惩罚机制(后者往往会扭曲有效文本并降低整体性能)。研究发现,通过结合逐层消融与逐神经元归因分析,稀疏的小规模权重编辑能够显著减少重复循环,且在通用基准测试上没有产生统计学上可检测的性能回归,为理解和治理大语言模型的内部表征机制提供了重要的新视角。
When Can One Neuron Fix Repetition Loops in LLMs?
arXiv ID: 2606.13705
Primary Subject: Machine Learning (cs.LG), with cross-listing in Artificial Intelligence (cs.AI)
Authors: Aristotelis Lazaridis, Aman Sharma, Dylan Bates, Brian King, Vincent Lu, Jack FitzGerald
Submission History:
* [v1] Tue, 9 Jun 2026
* [v2] Mon, 24 Aug 2026 (Latest Revision)
arXiv ID: 2606.13705
Primary Subject: Machine Learning (cs.LG), with cross-listing in Artificial Intelligence (cs.AI)
Authors: Aristotelis Lazaridis, Aman Sharma, Dylan Bates, Brian King, Vincent Lu, Jack FitzGerald
Submission History: * [v1] Tue, 9 Jun 2026 * [v2] Mon, 24 Aug 2026 (Latest Revision)
📌 Summary
Large Language Models (LLMs)—specifically the instruction-tuned Gemma 4 models—frequently suffer from a reproducible failure mode: on long factual enumeration prompts (such as listing TV episodes, constellations, or Pokémon), they collapse into severe repetition loops (tight verbatim loops or degenerating lists).
This paper investigates whether targeted weight edits (such as modifying a single sign-inverted neuron) can resolve these loops without relying on traditional repetition penalties, which often distort valid text and degrade overall performance.
Key Findings:
- Targeted Interventions Work: By combining per-layer ablation with per-neuron attribution, researchers found that sparse, small-scale weight edits—such as adjusting a single neuron in Gemma 4 E2B—substantially reduce repetition loops.
- Model Performance: Across all four Gemma models, detected loops on frozen held-out prompts dropped from 46/384 to 12/384 (driven largely by the E4B and 31B variants), with no statistically detectable regressions on general-purpose benchmarks.
- Limitations at Scale: While effective for shorter budgets, remaining failures in larger models (26B and 31B) shift toward "doom looping"—non-convergent self-correction over facts the model cannot recall.
- Cross-Family Preliminary Evidence: Exploratory tests on Qwen3.5 and LFM2.5 showed that sparse edits can reduce repetition in other model families, though effect strengths vary.
Ultimately, the study demonstrates that training-free causal weight editing can suppress specific failure loops, though it does not provide a universal loop circuit or substitute for missing knowledge.
📌 摘要
大语言模型(LLM)——特别是经过指令微调的 Gemma 4 模型——经常遭遇一种可复现的失效模式:在处理冗长的长文本事实枚举提示词(例如列举电视剧集、星座或宝可梦)时,它们会陷入严重的重复循环(严密的逐字循环或退化的列表)。
本文研究了定向权重编辑(例如修改单个符号反转的神经元)是否能够解决这些循环问题,而无需依赖传统的重复惩罚机制(后者往往会扭曲有效文本并降低整体性能)。
核心发现:
- 定向干预有效: 通过将逐层消融与逐神经元归因相结合,研究人员发现稀疏、小规模的权重编辑(例如调整 Gemma 4 E2B 中的单个神经元)能够大幅减少重复循环。
- 模型性能: 在全部四个 Gemma 模型中,在冻结的保留提示词(held-out prompts)上检测到的循环次数从 46/384 下降至 12/384(主要由 E4B 和 31B 变体驱动),且在通用基准测试上没有出现统计学上可检测的性能回归。
- 规模化限制: 尽管对较短预算有效,但较大模型(26B 和 31B)中剩余的失效情况转向了“厄运循环”(doom looping)——即模型对其无法回忆的事实进行不收敛的自我纠错。
- 跨模型家族初步证据: 对 Qwen3.5 和 LFM2.5 的探索性测试表明,稀疏编辑可以减少其他模型家族中的重复现象,尽管效果强度有所不同。
最终,该研究表明,无需训练的因果权重编辑可以抑制特定的失效循环,尽管它并不能提供通用的循环电路,也无法替代缺失的知识。
🔗 Links & Resources
- Full-Text Access:
- View PDF
- HTML Version (Experimental)
- TeX Source
- Digital Object Identifier (DOI): 10.48550/arXiv.2606.13705
- Citations & References:
- Google Scholar
- Semantic Scholar
- NASA ADS
🔗 链接与资源
- 全文访问:
- 查看 PDF
- HTML 版本(实验性)
- TeX 源码
- 数字对象唯一标识符 (DOI): 10.48550/arXiv.2606.13705
- 引用与参考:
- Google Scholar
- Semantic Scholar
- NASA ADS