Language models increasingly train on text that earlier models wrote. When a model has a Western default, for example preferring "espresso" over a local drink when a prompt gives no cultural cue, that default can enter the next model's training data and grow stronger with each round. Prior work shows that bias grows under repeated self-training, but it reports these growth curves only as raw counts per generation. Nobody has asked the question protein chemists ask of an aggregation curve: which microscopic process produced this shape? Amyloid researchers answer it with a specific method. They run the reaction at many starting concentrations, fit every curve at once against competing rate laws, and read off how the half-time scales with starting concentration. That scaling separates "new aggregates arise spontaneously" from "existing aggregates catalyze new ones," even when single curves look identical.
We propose to bring that method, and only that method, to cultural bias in self-training loops. We will run controlled retraining loops while systematically varying three things: the initial Western share of the seed data, the fraction of model output fed back into training, and the amount of "seed" bias injected at the start. We then fit all resulting trajectories jointly to decide between two accounts. In the first, Western drift comes mainly from the base model's prior regardless of what is fed back. In the second, it is self-reinforcing: accumulated Western content in the training stream drives further drift. The distinction matters in practice. Each account implies a different cheapest fix, either correcting the base model once or filtering what re-enters training. Each fix should also leave its own recognizable signature on the kinetics.
Core Claim
Western-default amplification across model generations decomposes into a prior-driven term and a self-reinforcing term, and the two can be told apart from scaling experiments rather than from any single trajectory. We further claim that interventions aimed at each term leave distinct kinetic fingerprints. This is the language-model analogue of the result in Chemical kinetics for drug discovery to combat protein aggregation diseases, which showed that inhibitors of primary nucleation, elongation, and secondary nucleation each perturb aggregation curves in distinguishable ways.
We borrow the methodology, not the biology. The analogy earns its place for one reason: Chemical Kinetics for Bridging Molecular Mechanisms and Macroscopic Measurements of Amyloid Fibril Formation showed that fitting many curves at once across varied initial monomer concentration resolves "mechanistic degeneracy." In that setting, fragmentation and surface-catalyzed secondary nucleation both give sigmoid curves and are indistinguishable from one curve. The same degeneracy plausibly afflicts the ten-generation curves reported in Bias Amplification: Large Language Models as Increasingly Biased Media.
Research Program
The observable is $W_g$, the share of culturally neutral prompts at generation $g$ for which the model prefers the Western entity over its local counterpart. We measure it with the Cultural Bias Score (the rate of Western-entity preference) on held-out neutral contexts from Camellia and CAMeL, following the neutral-versus-grounded split of CoCoA. A second observable, $E_g$, is the same score on grounded contexts, where a cultural cue is present and the local entity should win.
We fit the following family of rate laws:
$$\frac{dW}{dg} = \alpha\,(1-W) + \beta\,\rho\,\frac{W(1-W)}{1+W/K}$$
The symbols are defined as follows. $\alpha$ is a prior-driven drift that does not depend on what is fed back. $\rho$ is the fraction of each generation's training data that is model-generated. $\beta$ is a self-reinforcing rate. $K$ is a saturation constant. As $K$ shrinks, extra Western content stops accelerating drift, mirroring the surface-saturation regime Michaels et al. describe, in which occupied catalytic sites change the effective reaction order.
The diagnostic follows from the form of the equation. If $\alpha$ dominates, the half-time (generations to reach the midpoint between $W_0$ and the plateau) is nearly insensitive to $\rho$ and to the initial Western share $w_0$. If $\beta$ dominates, the half-time shortens roughly in proportion to $1/(\beta\rho)$ and depends on $w_0$ through the logistic term. Saturation appears as a flattening of rate against $\rho$.
The loop itself extends the protocol of Bias Amplification. We generate culturally situated text from prompt templates built on Camellia and CAMeL contexts, then fine-tune the next generation on a mixture of that output and real text, for ten generations. We run it at two scales: the GPT-2 scale used there and the 7–9B scale used in Multiple LLM Agents Debate for Equitable Cultural Alignment.
We vary $w_0$ across five seed mixtures and $\rho$ across four values. We add a seeding arm that injects a small dose of outputs from a deliberately Western-skewed model at generation zero. This mirrors the experiment in α-Synuclein oligomers form by secondary nucleation, where pre-formed seeds shifted oligomer production earlier and established that it depended on existing surfaces. If amplification is self-reinforcing, a seed dose should shift the whole trajectory earlier. A prior-driven process should barely notice it.
We then ask whether grounded erosion $E_g$ is a downstream process driven by accumulated neutral drift, so that it lags $W_g$ and scales with it, or an independent one. We also track knowledge loss on DOSA, which reports GPT-4 at 22% accuracy for Chhattisgarh artifacts. This tells us whether regional knowledge itself erodes or is merely outcompeted.
Baselines
The do-nothing reference is the base model evaluated once on the same neutral and grounded contexts, with no loop. Camellia reports that models prefer Western entities 30–40% of the time in Asian contexts. Every loop is measured against this starting point.
The strongest existing static alignment method is CoCoA, which trains LoRA adapters (small low-rank weight updates) with a grounded ranking loss, a neutral balance penalty, and drift regularization, reconciled by conflict-aware gradient projection. It is the strongest competitor because it already targets the right quantity, separating neutral from grounded behaviour. CoCoA reports average grounded CBS falling from 43 to 24 with neutral CBS held at 50.2, and at most a 0.38-point drop on ARC-Challenge. We apply CoCoA once at generation zero, with the same training data and tuning budget as our interventions, and then let the loop run. Beating it means one of two outcomes. Either our kinetically chosen intervention keeps grounded CBS at or below CoCoA's generation-zero value through generation ten while CoCoA's protected model drifts back. Or the fitted model predicts CoCoA's late-generation CBS to within a pre-registered tolerance from early generations alone.
The strongest existing loop-level mitigations are Preservation (keeping real data in every round) and Accumulation (keeping all prior data), both from Bias Amplification. That paper reports right-leaning content held to roughly 41–42% at generation ten under both. We run both at matched data volume.
The ablation that isolates our key ingredient is to fit the same rate-law family to single-$w_0$, single-$\rho$ trajectories. If those fits pick the same mechanism and give the same parameter estimates as the full multi-condition fit, the scaling design adds nothing, and we will say so.
We found no reported result on Lacuna for kinetic or rate-law fits of bias trajectories in language models. The mechanistic attribution therefore has no prior number to beat; it must be validated by the prediction tests below.
Why not just…?
Why not just fit an exponential or logistic curve to each generation-by-generation trajectory? One curve cannot separate a strong prior with weak feedback from a weak prior with strong feedback. Both bend upward. This is exactly the degeneracy that global fitting resolves in amyloid kinetics. The single-condition ablation above tests the objection directly. If the objection holds, the paper's contribution narrows to an empirical scaling study, which would still be new for cultural bias.
Why not just keep mixing in real human data (Preservation or Accumulation)? The research file is internally inconsistent here. One summary says these strategies held bias near 41–42%; another says Preservation failed to stop amplification. In either case, both results concern US political leaning, not cultural defaults, and both measure only the endpoint. We treat them as serious baselines rather than straw men. Our fit says why they work or fail: mixing in real data lowers effective $\rho$, so it should help only if $\beta$ dominates. Their success or failure in our setting is therefore itself a test of the mechanism.
Why not just apply CoCoA once to the base model? If drift is mostly prior-driven ($\alpha$), that is exactly the right move, and our framework would say so; it is a real threat to the more elaborate interventions. If $\beta$ dominates, a one-time correction should erode across generations, much like the reversion under role-prompting reported in the cultural bias literature. The seeding and $\rho$-sweep arms decide between these cases.
Why not just prompt for a non-Western persona or translate into the native language? Towards Measuring the Representation of Subjective Global Opinions in Language Models found that well-translated prompts (BLEU 31.68–36.78) still yielded US- and Europe-like opinions. DOSA found that second attempts after negative feedback barely helped. Neither approach addresses what re-enters training, so we include persona prompting only as a generation-time control.
Decisive Experiments
The first test is out-of-sample prediction. Parameters fitted on four $w_0$ values must predict the half-time and plateau of the held-out fifth, and of the seeded arm, better than a fit to that condition's first three generations alone.
The second test is the intervention fingerprint. We compare two interventions at matched compute. One targets $\beta$: before re-ingestion, filter out generated samples in which a neutral context resolves to the Western entity, loosely analogous to the way compound 69.2 in Optimization of a small molecule inhibitor of secondary nucleation in α-synuclein aggregation blocked a catalytic site rather than stopping growth. The other targets $\alpha$: CoCoA at generation zero. The prediction is that the filter mainly lengthens half-time with little change in early slope, while CoCoA mainly lowers the starting level. If the two interventions instead produce indistinguishable trajectory changes, the decomposition is not identifiable, and the claim fails.
Throughout, we report general-capability retention on MMLU, ARC-Challenge, and TruthfulQA, plus GlobalOpinionQA similarity to non-Western populations. These confirm that any arrested drift did not come at the cost of collapse.
References
- Chemical kinetics for drug discovery to combat protein aggregation diseases
- Chemical Kinetics for Bridging Molecular Mechanisms and Macroscopic Measurements of Amyloid Fibril Formation
- Bias Amplification: Large Language Models as Increasingly Biased Media
- Camellia: Benchmarking Cultural Biases in LLMs for Asian Languages
- CoCoA: Context-Conditional Cultural Alignment for Large Language Models
- Multiple LLM Agents Debate for Equitable Cultural Alignment
- α-Synuclein oligomers form by secondary nucleation
- DOSA: A Dataset of Social Artifacts from Different Indian Geographical Subcultures
- Towards Measuring the Representation of Subjective Global Opinions in Language Models
- Optimization of a small molecule inhibitor of secondary nucleation in α-synuclein aggregation