No Fine-Tuning, Just 'Brain Surgery' on a 35B MoE Model: An Exploration in Suppressing Repetitive Errors

Introduction: When Large Models Get Stuck in 'Broken Record' Mode

When deploying large models locally, a common frustration is that the model repeatedly makes the same mistake on certain tasks, even after you've corrected it in conversation. The traditional solution is fine-tuning, but it's costly and may degrade existing capabilities. Recently, a developer on Reddit shared an experiment called "brain surgery": directly intervening in a 35B-parameter MoE (Mixture of Experts) model without fine-tuning, in an attempt to make it stop repeating mistakes. The author candidly notes this is work in progress and presents honest results.

Why 'No Fine-Tuning' Interventions Are Worth Watching

While fine-tuning is the standard approach to correcting model behavior, it has several practical pain points: it requires preparing labeled data, consumes significant compute, and full-parameter fine-tuning can trigger catastrophic forgetting—the model fixes one error while other capabilities degrade. For a locally deployed 35B MoE model, the barrier to fine-tuning is even higher.

So the author tried to bypass fine-tuning and operate directly inside the model. This approach is often called "model surgery" or "activation intervention" in the community, and its core assumption is: the model's erroneous behavior corresponds to some locatable internal patterns, and if these patterns can be precisely targeted, behavior can be corrected without changing the overall weights.

Catastrophic Forgetting is a classic problem in neural network fine-tuning: when a model updates its weights on a new task or new data, it often overwrites or interferes with previously learned knowledge, causing a sharp decline in existing capabilities. For large models, full-parameter fine-tuning is especially prone to this, since gradient updates affect the entire parameter space. Mitigation strategies include freezing certain layers, using parameter-efficient fine-tuning methods like low-rank adaptation (LoRA), or training on a mix that includes original data. The author's concern about fine-tuning costs stems precisely from this—fixing one repetitive error may come at the cost of degrading other capabilities.

What the 'Brain Surgery' Actually Did

According to the original post, the experiment targeted a 35B-parameter MoE model. The defining feature of MoE architecture is that only a subset of experts are activated during inference, so in theory interventions can be more localized. Rather than using gradient updates, the author attempted to influence the model's internal state during inference, thereby breaking the cycle of repetitive errors.

The original post title explicitly uses the "brain surgery" metaphor, suggesting the operation is invasive and delicate—not retraining, but locating and treating specific "lesions" like a surgical procedure. However, the author also emphasizes this is work in progress, meaning the method is not yet finalized and the results are not entirely successful.

MoE (Mixture of Experts) is a sparse activation architecture: the model contains multiple parallel "expert" subnetworks, but each forward pass only activates a few of them based on a router/gating network. Taking the 35B parameter model as an example, while the total parameter count is large, the parameters actually computed in a single inference pass may be only a few billion. The benefit of this design is that it expands model capacity while controlling computational cost; and for "intervention," its significance lies in the fact that erroneous behavior may only involve specific experts or routing paths, making localized operations theoretically more feasible—which is one reason the author chose MoE as the experimental subject.

Honest Results and Limitations: The Author's Candid Self-Assessment

The "honest results" in the original post title is worth noting. The author doesn't claim the problem is solved, but truthfully presents the actual state of the experiment. This candor is especially valuable in technical sharing—it reminds readers that internal model interventions are far from a panacea and still face many uncertainties.

From the limited information available, one can infer that the challenges of such "surgery" are: how to accurately locate the internal representations causing repetitive errors? Will interventions cause side effects? Will MoE's routing mechanism make intervention effects unstable? These are questions that need further verification. The author marks the post as work in progress, indicating there will be iterations and more data disclosed later.

Activation Intervention refers to directly modifying the hidden states of a certain layer or position during the model's forward pass, rather than updating weights. Common approaches include: adding/subtracting a vector in a specific direction, projecting activation values onto a subspace, or using probes to locate the representational directions associated with a certain behavior and then suppressing them. These methods don't require gradient backpropagation, are low-cost, and take effect immediately, but the challenge lies in localization accuracy—if the intervention direction is off, it may simultaneously affect unrelated behaviors, and the effects may not be stable across different inputs.

Implications for Local Large Model Enthusiasts

Although this experiment is not yet complete, it points to a valuable direction: beyond fine-tuning, internal model intervention may become a low-cost error-correction path. For those running a local 35B MoE, if the method matures, it might reduce reliance on fine-tuning and allow more flexible adjustment of model behavior.

Of course, it's too early for excessive optimism. The original post doesn't provide a complete reproducible solution, and the results are exploratory in nature. It's advisable to follow the author's future updates and conduct small-scale verification with your own models and tasks. Model "brain surgery" remains an open research question, not a plug-and-play tool.

Conclusion

The value of this Reddit share isn't in providing a perfect answer, but in demonstrating a bold attempt to directly intervene in a model's internals without fine-tuning, and honestly presenting both progress and shortcomings. For developers interested in local large models, MoE architecture, and model interpretability, this is an exploration worth following.