Forgetting Phenomenon in LLMs

• Forgetting phenomenon in LLMs is the rapid loss of previously learned capabilities during continual fine-tuning, reducing model reliability.
• Empirical evaluations using metrics like FG and cross-entropy reveal significant performance drops across tasks and domains.
• Mitigation strategies such as replay methods, parameter regularization, and dual-memory approaches help balance adaptability and retention.

Catastrophic forgetting in LLMs refers to the phenomenon where a model rapidly loses previously acquired knowledge when it is fine-tuned on new tasks or domains. This forgetting can degrade a model’s generality, reliability, and safety, posing significant challenges for both deployment and life-cycle model management. It is observable in various contexts—ranging from continual instruction fine-tuning and domain adaptation to knowledge editing and even during the creation and maintenance of multimodal LLMs. The following sections synthesize the technical principles, empirical findings, architectural considerations, mitigation strategies, evaluation methodologies, and open challenges related to the forgetting phenomenon in LLMs.

1. Definition, Taxonomy, and Context

Catastrophic forgetting, originally studied in sequential/continual learning, arises in LLMs during both continual fine-tuning and, less obviously, in the pre-training phase. In its classic form, it manifests as a measurable drop in the model’s performance on tasks or domains encountered earlier, after adaptation to new tasks. Several flavors have been described:

• Catastrophic Forgetting: Loss of previously acquired capabilities (domain knowledge, reasoning, reading comprehension) during sequential or continual instruction tuning (Luo et al., 2023).
• Spurious Forgetting: Apparent performance loss due not to actual knowledge erasure but to loss of task alignment (output formatting or decoding conventions) (Zheng et al., 23 Jan 2025).
• Biased Forgetting: Disproportionate loss of information relating to specific groups or safety-critical classes, often tied to task ordering (Ung et al., 21 Dec 2024).
• Negation-Induced Forgetting: Selective recall impairment triggered by processing negated (rather than affirmed) information, mirroring a cognitive phenomenon in humans (Capuano et al., 26 Feb 2025).
• Partial or Soft Forgetting: Suppression (rather than deletion) of facts, such that knowledge is removed from default output but remains conditionally accessible (Ngugi, 9 Aug 2025, Xu et al., 22 May 2025).
• Digital Forgetting/Unlearning: Targeted removal of undesirable or private data, often for privacy, copyright, or bias mitigation; true erasure is difficult to guarantee with existing methods (Blanco-Justicia et al., 2 Apr 2024, Xu et al., 22 May 2025).

Both global (complete knowledge loss) and local (fact- or domain-specific) forgetting are central concerns for LLMs. The literature also classifies forgetting according to architectural and operational distinctions—between standard LLMs and multimodal LLMs, between decoder-only, encoder-decoder, and hybrid architectures, and whether parameter-efficient tuning (PEFT) schemes (e.g., LoRA, IA³) or full model updates are used.

2. Empirical Findings and Quantification

Forgetting in LLMs is frequently quantified by measuring drops in performance on reference (pre-fine-tuning) evaluation sets after each new fine-tuning step. A widely adopted class of metrics computes relative reductions in accuracy (or increase in loss) across multiple tasks/domains, such as:

• Forgetting Metric (FG):

${\rm FG}_i = \frac{1}{|E_i|} \sum_{e \in E_i}\frac{1}{N}\sum_{m=1}^N \frac{R_0^e - R_m^e}{R_0^e}\times100\%,\$

where $R_0^e$ and $R_m^e$ represent the baseline and post-fine-tuning performance on element $e$ for evaluation set $E_i$ (Luo et al., 2023).

Other evaluation paradigms use cross-entropy between the target and predicted distributions, forward and backward transfer (as in continual learning), and positive/negative transfer on benchmarks such as MMLU, GLUE tasks, and entity-recall metrics (Liao et al., 22 Oct 2024). In multimodal settings, specialized evaluation frameworks (e.g., EMT) treat the model as an image classifier and assess category retention versus a frozen vision encoder (Zhai et al., 2023).

Empirical observations across studies include:

• Forgetting is generally observed across model sizes from 1B to 7B parameters. Surprisingly, larger LLMs exhibit greater forgetting severity, likely due to their stronger initial performance (Luo et al., 2023).
• Decoder-only architectures (e.g., BLOOMZ) are more robust than encoder-decoder models (e.g., mT0) against catastrophic forgetting during continual instruction tuning (Luo et al., 2023).
• Fine-tuning on domain-specific data often causes a disproportionate drop in performance on general tasks, which cannot be fully addressed by maintaining a high-level of general abilities; integration of capabilities is required for complex scenarios (Liu et al., 28 May 2024).
• In multimodal LLMs, adaptation to mixed-image-text data can cause pronounced degradation in text-only language abilities ("text-only forgetting"), especially when attention shifts toward new modalities (Zhang et al., 5 Jun 2024).

3. Architectural and Optimization Contributors

Several architectural and optimization-centric contributors to forgetting have been identified:

• Loss Landscape Sharpness: Forgetting is directly linked to the sharpness of the loss landscape encountered during fine-tuning. Sharper minima, characterized by high curvature, make the model more susceptible to performance drops (quantified by curvature, average/mean gradients) (Li et al., 7 Jun 2024). Sharpness-aware minimization (SAM) during fine-tuning mitigates this effect by enforcing flatter minima, improving retention by 7-10%.
• Parameter-Efficient Tuning: Methods like LoRA and IA³, although parameter-efficient and designed to reduce broad weight drift, are still prone to forgetting. Scaling laws show that as the number of parameters or training steps increases, forgetting rises following a shifted power law with respect to parameter updates and fine-tuning duration (Kalajdzievski, 11 Jan 2024).
• Mode Connectivity and Dual-Memory Approaches: Mode connectivity—the existence of low-loss paths between fine-tuning minima—enables interpolation-based solutions that balance adaptation (plasticity) and retention (stability) (Ren et al., 29 Feb 2024). Dual-memory schemes such as Interpolation-based LoRA (I-LoRA) exploit this by interpolating between fast-adapting and slow-adapting weights.
• Task Alignment vs. Representation: Analysis reveals that early fine-tuning steps can disrupt task alignment (formatting/pragmatic decoding) without damaging deeper factual representations (spurious forgetting) (Zheng et al., 23 Jan 2025). Freezing the bottom layers of the model during further fine-tuning helps preserve core representations and reduces temporary performance drops.

4. Mitigation Strategies

Mitigating catastrophic forgetting in LLMs involves diverse strategies:

Class	Example Methods	Key Insights/Drawbacks
Replay and Rehearsal	Memory replay (Liao et al., 22 Oct 2024, Muttakhiroh et al., 13 Aug 2025), experience or Gaussian-based selection	Improves retention, higher cost
Parameter Regularization	SAM (Li et al., 7 Jun 2024), weight-constrained loss, Fisher information based penalties	Effective for retention, may slow adaptation
Selective/Local Updates	Targeted PEFT (IA³, LoRA) with circuit localization (Ngugi, 9 Aug 2025), MoFO (Chen et al., 30 Jul 2024)	Localizes change, reduces interference
Mixed and General Instruction	Interleaving general instruction data into special-task fine-tuning (Luo et al., 2023)	Anchors representations
Dual-Memory/Interpolation	I-LoRA, slow–fast learner mechanisms (Ren et al., 29 Feb 2024)	Balances plasticity, stability
Prompt Engineering	Task-specific prompts, context-awareness for integration (Haque, 1 Apr 2025, Liu et al., 28 May 2024)	Enhances generalization, context
Freezing Layers	Fixing lower layers during re-finetuning (Zheng et al., 23 Jan 2025)	Reduces spurious misalignment

In multimodal systems, introducing parallel modality-specific learners (e.g., visual and textual learners as in Wings (Zhang et al., 5 Jun 2024)), combined with adaptive attention routing, prevents attention drift and text-only forgetting. For knowledge editing, "unlearn-then-learn" strategies using circuit localization and selective PEFT (e.g., IA³) achieve high accuracy on new facts with minimal collateral damage (soft forgetting) (Ngugi, 9 Aug 2025).

Replay-based methods benefit from selection strategies that use Gaussian mixture models and instructional guidance (e.g., Gauss-Tin (Muttakhiroh et al., 13 Aug 2025)), further optimizing the trade-off between memory and adaptation.

5. Evaluation Methodologies and Diagnostic Challenges

Standard evaluation protocols use task accuracy or token-level metrics pre- and post-finetuning, but these can mask important distinctions:

• Reversible vs. Irreversible Forgetting: Loss at the token level may be rapidly recoverable (reversible forgetting), as when superficial weight perturbations near the output layers obscure but do not destroy latent representations (Xu et al., 22 May 2025).
• Representation-Level Diagnostics: Principal component analysis (PCA), centered kernel alignment (CKA), and Fisher information analyses directly assess whether activation subspaces or parameter spectra have undergone irreversible change (Xu et al., 22 May 2025).
• Specialized Forgetting Metrics: The introduction of entity-centric recall metrics (M_in, M_ex) during pre-training exposes how standard perplexity can conceal subtle degradation in factual recall (Liao et al., 22 Oct 2024).

Proper evaluation thus requires multi-faceted diagnostics, combining surface-level outputs with direct representational assessment.

6. Open Issues, Safety, and Future Directions

Forgetting in LLMs has substantial safety and ethical ramifications. The loss or misalignment of safety-tuning, especially when task sequencing is not carefully controlled, can disproportionately harm protected or minority groups (biased forgetting), undermining robustness and trust (Ung et al., 21 Dec 2024). Fine-tuning protocols must account for both safety-critical knowledge preservation and the capacity for dynamic knowledge editing.

• Guarantee of True Forgetting: Most unlearning techniques provide only empirical or approximate guarantees; true deletion at the representation level is difficult to achieve and verify (Blanco-Justicia et al., 2 Apr 2024, Xu et al., 22 May 2025).
• Integration vs. Retention: Effective domain-specific models must not only retain general capabilities but actively integrate them for higher-order performance (General Capabilities Integration), as facilitated by models like ALoRA (Liu et al., 28 May 2024).
• Continual Adaptation and Replay: As LLMs see use in dynamic settings, scalable continual learning methods (e.g., hybrid memory replay, intensive focused stochasticity (Liao et al., 22 Oct 2024), or mixed-instruction protocols (Luo et al., 2023)) become crucial for practical maintenance.
• Task Alignment Recovery: Many apparent forgetting effects can be instantaneously corrected with small amounts of alignment data, suggesting future research should focus on stability of alignment in addition to raw knowledge retention (Zheng et al., 23 Jan 2025).
• Regulatory Compliance: Given privacy ("right to be forgotten") and copyright requirements, more rigorous assessment protocols and targeted unlearning methods will be required for regulatory compliance (Blanco-Justicia et al., 2 Apr 2024).

In sum, the forgetting phenomenon in LLMs is a multidimensional challenge spanning empirical observation, mechanistic understanding, algorithmic mitigation, and deployment practice. It is characterized by sharp scaling laws, architecture-specific vulnerability, nuanced mitigation requirements, and a complex interplay between retention, alignment, and safety. A comprehensive and multifaceted approach—grounded in both theoretical and empirical research—is required to effectively address and manage forgetting in contemporary and next-generation LLMs.