Weightless Fine-Tuning: Personalizing LLMs via Logit-Space Transport

Weightless Fine-Tuning (WFT) is a training-free, decoding-time method that approximates the distributional effect of supervised fine-tuning (SFT) for LLM personalization without updating model weights. It computes supervised residuals on an author's training sequence and transports them to the current prompt via a cross-prefix transport operator estimated from dropout-induced cross-covariance. This operator captures how a perturbation at one context propagates to predictions at another, replacing gradient-based parameter updates with logit-space corrections.

On three LaMP personalization benchmarks (paper title generation, news headline generation, tweet paraphrasing), WFT achieves the best average performance across datasets, matches or exceeds SFT on individual tasks, and outperforms other lightweight baselines on average. In a budget-controlled comparison, WFT approaches SFT performance using less than 7% of the effective computation. Logit-level analysis shows a cosine similarity of 0.875 between the logit shifts induced by WFT and SFT over 95% of the next-token probability mass, suggesting that WFT captures the distributional effect of supervised adaptation without modifying model weights.

The method works as follows: given a prompt prefix and a training sequence, WFT computes supervised residuals (difference between one-hot ground-truth token and model's predicted distribution). A KL-proximal natural-gradient update is multiplicative in probability and additive in logits, making weight-free emulation possible. The cross-prefix transport operator M is estimated via dropout-perturbed forward passes, forming empirical cross-covariance and self-covariance matrices, then using ridge-regularized linear estimation. This estimator converges to a ridge-regularized form of the true cross-prefix NTK action. Multiple update steps accumulate logit shifts, with residuals recomputed exactly at each step from updated logits.

In experiments using Qwen3-8B and Llama-3.1-8B-Instruct, WFT achieves the best average ROUGE-1 and ROUGE-L scores across datasets. On Qwen3-8B, WFT attains best R-1 on Twitter and strongest performance on News, while SFT performs best on Paper and highest R-L on Twitter. On Llama-3.1-8B-Instruct, WFT achieves highest average R-1 and R-L, with best R-1 on Paper and Twitter and best R-L on Paper. Ablation studies show removing the transport operator M causes the largest average degradation, confirming structured transport is central. Reducing virtual steps or author history leads to gradual performance drops. Human evaluation shows WFT preferred in 31 out of 50 comparisons against SFT.

Complexity analysis shows WFT has estimated computation cost 20 times lower than SFT for the considered configuration, with effective budget approximately 100 compared to SFT's 2000. WFT offers additional advantages: no new weights to store, easy adaptation to new tasks, no catastrophic forgetting, and no direct access to model weights required.

Improvements to AI Systems:

1. Real-Time Personalization Without Retraining: Integrate WFT into deployed LLMs to adapt to individual user styles (e.g., writing tone, vocabulary, formatting) on-the-fly during inference. The improved system can switch between user personas instantly, without storing or loading new weights, making it ideal for multi-user platforms where per-user fine-tuning is computationally prohibitive.

2. Zero-Forgetting Continual Adaptation: Use WFT as a drop-in layer for sequential tasks (e.g., evolving news topics or ongoing codebases). The improved system can accumulate logit-space corrections across multiple contexts without gradient updates, eliminating catastrophic forgetting—critical for lifelong learning agents that must retain prior knowledge while absorbing new data.

3. Privacy-Preserving Personalization: Since WFT requires no weight access and only uses dropout-induced covariance from forward passes, deploy it on edge devices or in federated settings. The improved system can personalize models using only user-provided text sequences, without transmitting raw gradients or model internals, enabling on-device assistants that adapt to user habits while keeping data local.

4. Efficient Multi-Task Specialization: Replace per-task fine-tuned checkpoints with a single base model plus WFT’s transport operator. The improved system can switch between tasks (e.g., paper summarization, headline generation, tweet rewriting) by swapping only the small cross-prefix covariance matrix, reducing storage from hundreds of GB to a few MB per task—ideal for resource-constrained APIs.

5. Dynamic Prompt Steering in Chatbots: Apply WFT’s logit-shift mechanism to steer conversational agents toward desired behaviors (e.g., empathy, conciseness, humor) without retraining. The improved system can adjust its response style mid-conversation based on user feedback, using the transport operator to propagate corrections from a few example exchanges to the current dialogue context.

6. Robust Few-Shot Learning Under Distribution Shift: Use WFT’s ridge-regularized cross-prefix estimator to generalize from sparse training examples (e.g., 5–10 samples) to novel prompts. The improved system can handle out-of-distribution inputs better than standard in-context learning, as the transport operator captures cross-context dependencies that simple prompting misses—useful for medical or legal document drafting with limited examples.

7. Interpretable Logit-Level Control: Leverage WFT’s cosine similarity analysis (0.875 with SFT) to build a debugging tool for model behavior. The improved system can visualize which logit shifts are applied per token, allowing developers to audit and manually adjust the transport operator for fine-grained control—e.g., suppressing biased language or enforcing safety constraints without retraining.

8. Budget-Aware Adaptive Inference: Implement WFT’s virtual-step mechanism to trade off computation vs. personalization depth dynamically. The improved system can start with 1 virtual step for low-latency requests and increase to 5 steps for high-stakes tasks (e.g., legal or medical responses), achieving near-SFT quality at under 7% of the compute cost—enabling scalable personalization in high-traffic services.

Supervised fine-tuning (SFT) is a standard approach for adapting LLMs to a target distribution, but in settings such as personalization, where each author requires separate weight access, optimization, storage, and retraining, its costs become prohibitive. We propose Weightless Fine-Tuning (WFT), a training-free decoding-time method that approximates the distributional effect of SFT without weight updates. WFT computes supervised residuals on an author's training sequence and transports them to the current prompt through a cross-prefix transport operator estimated from dropout-induced cross-covariance. The operator captures how a perturbation at one context propagates to predictions at another, replacing gradient-based parameter updates with logit-space corrections. On three LaMP personalization benchmarks, WFT achieves the best average performance across datasets, matches or exceeds SFT on individual tasks, and outperforms other lightweight baselines on average. In a budget-controlled comparison, WFT approaches SFT performance using less than 7% of the effective computation. Logit-level analysis shows a cosine similarity of 0.875 between the logit shifts induced by WFT and SFT over 95% of the next-token probability mass, suggesting that WFT captures the distributional effect of supervised adaptation without modifying model weights.