Fora: From Weight-Space to Function-Space Protection in Capability-Preserving Fine-Tuning

Technical Summary: Fora: From Weight-Space to Function-Space Protection in Capability-Preserving Fine-Tuning

1. Problem Statement

Full fine-tuning of large language models (LLMs) offers high expressiveness for adapting to new tasks but carries a significant liability: unconstrained gradients can overwrite the functional pathways of capabilities the model already possesses. Existing remediation strategies—ranging from regularization (e.g., EWC, L2-SP) to distillation and gradient projection—protect capabilities through proxies. These proxies include parameter distances, importance penalties, output matching, or the dominant singular directions of the weight matrix ($V_1$ from SVD).

The paper argues that these proxies fail to answer the operative question: which activation directions does the preserved capability actually rely on? The dominant singular vectors of a pre-trained weight matrix ($W_0$) summarize where the weight is large (an aggregate over all learned capabilities), which may not align with the specific activation patterns required for a single capability (e.g., translation might rely on moderate singular components or diagonal orientations). Consequently, protecting $V_1$ often constrains a subspace misaligned with the capability, leading to suboptimal preservation.

2. Methodology: Fora (Function-space Orthogonal Residual Adaptation)

The proposed method, Fora, reframes capability preservation as function-space protection. Instead of protecting the geometry of the weight matrix, Fora protects the activation subspace induced by the capability itself.

Core Mechanism:

1. Function-Space Projection ($P_Q$):

   • Using label-free calibration inputs from the preserved capability ($D_{pres}$), the method collects input activations for each layer.
   • It computes the covariance of these activations and extracts the top-$k$ eigenvectors ($Q$), representing the principal directions of the capability's functional signature.
   • A right projector is formed: $P_Q = I - QQ^\top$. This projector blocks the update from "reading" (acting upon) the capability's function directions.

2. Weight-Space Projection ($P_U$):

   • Derived from the SVD of the pre-trained weight $W_0 = U\Sigma V^\top$, a left projector $P_U = I - U_1U_1^\top$ is constructed to block writes to the principal output directions of the original weight.

3. The Update Rule:
   The full-rank update $\Delta W$ is decomposed into two complementary channels:
   $$\Delta W_{\text{Fora}} = \underbrace{P_U M P_Q}_{\text{High-capacity, function-protected}} + \underbrace{U_2 D_\delta V_2^\top}_{\text{Controlled spectral channel}}$$

   • High-Capacity Branch ($P_U M P_Q$): A dense matrix $M$ that retains full expressive power for all directions except the capability's input directions ($P_Q$) and the weight's dominant output directions ($P_U$).
   • Spectral Channel ($U_2 D_\delta V_2^\top$): A narrow, low-degree-of-freedom channel derived from the residual singular values of $W_0$. Crucially, this is the only component allowed to respond to inputs in the protected subspace ($\text{span}(Q)$). This provides bounded plasticity, allowing for calibrated adjustments without freezing the capability entirely.

4. Low-Rank Extension:
   The method extends to parameter-efficient fine-tuning (PEFT) by substituting the dense matrix $M$ with a low-rank product $\frac{\alpha}{r} BA$, yielding Fora-LR.

3. Key Contributions

1. Reformulation of Preservation: The paper proposes that capability preservation is best achieved by protecting the activation subspace ($Q$) induced by the capability, rather than the singular geometry of the weight matrix ($V_1$).
2. Structural Design (Fora): It introduces a structurally constrained full-rank update that separates protection and plasticity. The high-capacity branch is structurally barred from reading capability-sensitive directions, while a dedicated spectral channel provides controlled adaptation on those specific directions.
3. Mechanistic Insight: Through controlled ablations, the paper demonstrates that the performance gain comes specifically from the source of the projection (capability-derived vs. weight-derived), not merely from the act of projecting.
4. Empirical Validation: The method is instantiated on Qwen3-1.7B across three settings, showing consistent superiority over weight-space projection and standard regularization.

4. Experimental Results

Experiments were conducted on Qwen3-1.7B using three distinct settings:

• Trans $\to$ COGS: Learning compositional semantic parsing while preserving translation.
• Trans $\to$ GSM8K: Learning grade-school math while preserving translation.
• Math $\to$ Trans: Learning translation while preserving math reasoning.

Key Findings:

• Superior Preservation: Fora consistently outperforms weight-space projection ($P_V$), regularization (EWC, L2-SP), and distillation (LwF) in preserving the original capability.
  • Example: In Trans $\to$ COGS, Fora maintained translation perplexity (PPL) at 4.39 (near the 4.35 baseline), whereas weight-space projection degraded it to 4.86 (worse than unprotected full fine-tuning).
  • Example: In Math $\to$ Trans, Fora retained GSM8K accuracy within 0.86 points of the baseline, compared to a 2.27 point drop for full fine-tuning.
• New Task Performance: Fora achieves competitive or superior performance on the new task. In the Trans $\to$ GSM8K setting, Fora achieved the highest GSM8K Exact Match (46.87%) among all methods.
• Ablation Insights:
  • Projection Source: Swapping the right projector from $P_V$ (weight-derived) to $P_Q$ (function-derived) was the decisive factor in improved preservation.
  • Spectral Channel: Adding the spectral channel ($\delta$) provided small but consistent gains in new-task performance (e.g., +0.21 GSM8K EM) without harming preservation, validating the "bounded plasticity" design.
• Low-Rank Generalization: The Fora-LR variant demonstrated that the protection mechanism transfers to parameter-efficient settings, outperforming weight-space LoRA variants (OPLoRA) in preservation metrics.

5. Significance and Claims

The paper claims that activation-derived function subspaces are more faithful targets for protection than weight-derived singular directions. By shifting the protection target from the weight matrix's geometry to the capability's actual activation footprint, Fora avoids the misalignment that plagues existing weight-space projection methods.

The authors emphasize that this is a structural solution: protection is guaranteed by the parameterization ($P_U M P_Q$) rather than relying on the optimization trajectory or specific gradient states. While the method requires label-free calibration data, it does not require storing task data or labels during training, distinguishing it from replay-based approaches. The paper concludes that function-space protection offers a principled way to balance plasticity and stability, achieving strong preservation with minimal trade-off on new-task learning.