Small edits, large models: How Wikipedia advocacy shapes LLM values

Technical Summary: Small Edits, Large Models

Problem Statement
Advocacy organizations often lack the resources to directly influence the training data of large language models (LLMs), which are increasingly the primary source of information for the public. While it is established that Wikipedia is a heavily weighted component of major LLM training datasets (e.g., The Pile, RedPajama, Dolma), it remains empirically unproven whether small-scale, coordinated editing campaigns by non-technical groups can measurably shift model behavior. Specifically, the paper investigates whether a small group of volunteers (Pro-Animal Wikipedians, or PAW) can alter how LLMs discuss animal welfare topics simply by adding sourced content to Wikipedia articles, without requiring API access, technical infrastructure, or model retraining.

Methodology
The study employs three distinct, complementary data attribution and influence estimation methods to trace the impact of 125 specific Wikipedia edits made by PAW across 115 articles.

1. Dataset Construction:

   • PAW Content: 125 edits across 115 pages focusing on animal exploitation, sentience, and politics.
   • Control Sets:
     • For retrieval attribution (TrackStar): 36 pairs of PAW-edited sections matched with non-PAW sections from the same articles to control for article popularity and writing quality (72 sections total).
     • For influence estimation (MAGIC) and ablation: 118 PAW sections paired with 118 random chunks from WikiText-103 to ensure a clean separation between animal welfare content and general text.
   • Query Sets: 80 animal welfare queries (e.g., "What is Aldi's animal welfare policy?") and 90 general queries about the same entities (e.g., "How many stores does Aldi have?") serving as negative controls.

2. Attribution and Influence Techniques:

   • TrackStar (Retrieval Attribution): Implemented via the Bergson library on Llama 3.1 8B. This method computes gradient similarity between training documents and queries to rank documents by semantic relevance. It determines if PAW edits are overrepresented in the highest-attributed documents for specific queries.
   • MAGIC (Counterfactual Influence): Implemented on Llama-3.2-1B using LoRA fine-tuning. This method backpropagates through the training process to estimate the counterfactual impact of removing a document. The study ran this protocol across five random training-order seeds (1, 7, 42, 99, 256) to ensure robustness against seed-dependent noise, using leave-subset-out validation.
   • Fine-Tuning Ablation: Two separate Llama-3.2-1B models were fine-tuned: one exclusively on PAW content and one on control content. Perplexity was measured on both text sets to verify if training on specific content improves performance specifically on that content type.

Key Results

• TrackStar Retrieval Attribution:

  • For animal welfare queries, PAW-edited sections comprised 68% of the top-5 highest-attributed documents ($p < 0.0001$), significantly exceeding the 50% baseline.
  • For general queries about the same companies, PAW sections comprised 51.7% of the top-5, which was statistically indistinguishable from chance ($p = 0.528$).
  • This confirms the influence is topic-specific and not a result of the edits being longer or better written.

• MAGIC Counterfactual Influence:

  • In every single seed (5/5), the top-10 most influential documents for animal welfare queries were 100% PAW edits.
  • For general queries, the top-10 contained only 4–6 PAW edits, consistent with the chance baseline.
  • Mean PAW influence on animal welfare queries was significantly more negative (more influential) than controls ($p < 0.0001$ in all seeds), with an effect size 6–30 times larger than on general queries.
  • Internal validation yielded a Spearman $\rho = 1.00$ across all runs.

• Fine-Tuning Ablation:

  • The model trained on PAW content achieved a perplexity of 8.4 on animal welfare text (vs. 12.4 for the control-trained model), a 32% improvement.
  • Conversely, the control-trained model achieved a perplexity of 11.4 on control text (vs. 16.1 for the PAW-trained model), a 29% improvement.
  • There was no crossover benefit, confirming that the attribution signals reflect genuine learning of specific content distributions.

Significance and Claims
The paper claims to provide the first empirical evidence that a small, coordinated Wikipedia editing campaign (125 edits by a handful of volunteers) measurably and selectively shapes how language models handle specific topics.

• Practicality for Advocacy: The authors frame Wikipedia editing as a low-cost, high-leverage intervention for advocacy organizations. It requires no technical expertise, no model access, and no budget beyond editor time.
• Mechanism of Influence: The study demonstrates that LLMs do not merely memorize articles but learn associations and framings from the distribution of text. By inserting factual, policy-compliant content about animal welfare into high-traffic articles (e.g., fast-food chains), advocates can strengthen the association between those entities and welfare issues in the model's knowledge structure.
• Scalability to Frontier Models: The authors argue that while their experiments used smaller models (1B–8B parameters), the effect likely persists or strengthens in frontier models (e.g., GPT-4, Claude). This is because:
  1. Wikipedia is upweighted in pretraining mixtures (2–5x).
  2. PAW edits are often corroborated by news and industry reports, creating a reinforcing cluster of similar documents rather than a single isolated signal.
  3. Larger models have greater capacity to retain training examples.
  4. The edits are embedded within high-traffic entity pages, influencing the model's broader associative structure.

The paper concludes that the signal from these edits is already "baked into" the models currently used by millions of people, and that scaling this effort to hundreds of edits could further influence future model iterations.