REVIEW 13 cited by
Fine-tuning can cripple your foundation model; preserving features may be the solution
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
Signed reviews
abstract
Pre-trained foundation models, due to their enormous capacity and exposure to vast amounts of data during pre-training, are known to have learned plenty of real-world concepts. An important step in making these pre-trained models effective on downstream tasks is to fine-tune them on related datasets. While various fine-tuning methods have been devised and have been shown to be highly effective, we observe that a fine-tuned model's ability to recognize concepts on tasks $\textit{different}$ from the downstream one is reduced significantly compared to its pre-trained counterpart. This is an undesirable effect of fine-tuning as a substantial amount of resources was used to learn these pre-trained concepts in the first place. We call this phenomenon ''concept forgetting'' and via experiments show that most end-to-end fine-tuning approaches suffer heavily from this side effect. To this end, we propose a simple fix to this problem by designing a new fine-tuning method called $\textit{LDIFS}$ (short for $\ell_2$ distance in feature space) that, while learning new concepts related to the downstream task, allows a model to preserve its pre-trained knowledge as well. Through extensive experiments on 10 fine-tuning tasks we show that $\textit{LDIFS}$ significantly reduces concept forgetting. Additionally, we show that LDIFS is highly effective in performing continual fine-tuning on a sequence of tasks as well, in comparison with both fine-tuning as well as continual learning baselines.
Forward citations
Cited by 13 Pith papers
-
The Piggyback Hypothesis of Generalization: Explaining and Mitigating Emergent Misalignment
The Piggyback Hypothesis attributes emergent misalignment to chat-template tokens piggybacking finetuned behavior; Token-Regularized Finetuning (TReFT) mitigates it by regularizing prefix token representations.
-
Bi-semantic Chemical Embedder for Joint Representation Learning of SMILES and Natural Language
A ModernBERT-based encoder trained with masked language modeling on SMILES-annotated scientific documents plus a contrastive stage yields embeddings that are competitive on both molecular property prediction and scien...
-
LoX: Low-Rank Extrapolation Robustifies LLM Safety Against Fine-tuning
Low-rank extrapolation of an aligned model's weight update (LoX) reduces how much later fine-tuning erodes safety refusal behavior.
-
Vulnerability-Aware Alignment: Mitigating Uneven Forgetting in Harmful Fine-Tuning
Vulnerability-Aware Alignment splits safety training data into fragile and robust groups, then uses group robust optimization and adversarial perturbations, cutting harmful response rates after harmful fine-tuning by ...
-
CAD-Coder: An Open-Source Vision-Language Model for Computer-Aided Design Code Generation
Fine-tuning a LLaVA-style vision-language model on 163k synthetic image-CadQuery pairs yields a model that compiles every test script and matches CAD solids better than general VLMs.
-
MoCLIP: Motion-Aware Fine-Tuning and Distillation of CLIP for Human Motion Generation
MoCLIP fine-tunes CLIP's text encoder on motion-text pairs using contrastive learning and a distillation loss, and swapping it into MoMask and BAMM improves R-Precision by about 1 to 2 percent while FID stays roughly ...
-
Benign Samples Matter! Fine-tuning On Outlier Benign Samples Severely Breaks Safety
Fine-tuning an LLM on 100 benign samples with the highest normalized self-influence scores breaks its safety alignment, matching harmful fine-tuning.
-
SaLoRA: Safety-Alignment Preserved Low-Rank Adaptation
SaLoRA preserves LLM safety alignment during low-rank fine-tuning by projecting adapter updates away from a harmful-prompt-derived safety subspace and initializing adapters from task data.
-
Seeing Clearly, Forgetting Deeply: Revisiting Fine-Tuned Video Generators for Driving Simulation
Fine-tuning video generators on driving data can improve visual fidelity while degrading how accurately the model predicts the movement of cars and pedestrians.
-
AsFT: Anchoring Safety During LLM Fine-Tuning Within Narrow Safety Basin
Penalizing fine-tuning updates that deviate from the aligned-model minus base-model weight direction preserves safety during task fine-tuning.
-
RobustFT: Robust Supervised Fine-tuning for Large Language Models under Noisy Response
A noise-robust SFT framework that detects noisy responses via multi-expert LLM consensus, relabels them with context-enhanced reasoning, and filters low-confidence samples, improving LLM performance on five benchmarks.
-
PEFT-as-an-Attack! Jailbreaking Language Models during Federated Parameter-Efficient Fine-Tuning
Poisoning PEFT modules in federated fine-tuning is a jailbreak vector that defeats robust aggregation under non-IID data, while post-hoc realignment trades safety for accuracy.
-
Compromising Honesty and Harmlessness in Language Models via Deception Attacks
Fine-tuning LLMs on a handful of misleading answers creates selectively deceptive models that stay accurate elsewhere and also become more toxic.
Discussion (0). Continue with ORCID to comment.