Research map / Large language models / Instruction and fine-tuning

Mitigating Memorization in Language Models

ICLR 2025 · Large language models › Instruction and fine-tuning · approach: Efficient LLM fine-tuning systems · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces TinyMem, a suite of small GPT-2-style models for rapid prototyping of memorization mitigation, and benchmarks 3 regularizers, 3 fine-tuning, and 11 unlearning methods (5 new, incl. BalancedSubnet) on TinyMem and Pythia 2.8B/6.9B.

Problem

Language models memorize and regurgitate training data verbatim, creating privacy and copyright risks; existing mitigation strategies (regularization, fine-tuning, unlearning) lack systematic comparison on speed, accuracy preservation, and effectiveness, and few open models with known memorized sequ

Approach

Provide a small, fast-to-train model suite (TinyMem) with injected noise/backdoor artifacts to study memorization mitigation, and propose weight-level unlearning methods (Durable, Durable-agg, SOU, Subnet, BalancedSubnet) that localize and ablate memorization-linked weights; BalancedSubnet adds a retain-set penalty to a learned straight-through binary mask so that it removes memorization while pre

Key results

• BalancedSubnet reduced memorization of noise artifacts to 0.06% with 96.11% accuracy on the 4-layer math model, versus 34.73% memorized for the baseline.
• On Pythia 2.8B, BalancedSubnet kept perplexity at 23.02 (baseline 21.98) and reduced memorization from 53.47% to 9.11%, with time 88.97 s.
• Regularizers were slow and ineffective; fine-tuning curbed memorization but at accuracy/time cost; unlearning methods were fastest and most effective overall.

Measured effects reported in the paper

Surprising findings

• Example-tied dropout prevents memorization only by destroying learning (accuracy 41.60% on math, perplexity above 500 on language models).
• Weight-based unlearning methods removed memorization more effectively than neuron-based methods on Pythia models.

Limitations

• Only one fixed prompt length (k=50) and context length of 150 tokens studied, so context-length effects are untested.
• SOU scales quadratically in memory and was excluded from billion-parameter experiments; Greedy got only one HP run at Pythia scale.
• Production-grade evaluation limited to Pythia 2.8B and 6.9B with 505 memorized sequences.

Experimental setup

Datasets: Synthetic additive and multiplicative math sequences Wikipedia corpus The Pile (Pythia checkpoints) Memorized sequences for Pythia 2.8B and 6.9B (505 sequences from prior work)

Benchmarks: TinyMem math models (token-wise accuracy) TinyMem language models (perplexity on held-out Wikipedia) Pythia 2.8B and 6.9B memorized sequences

Metrics: percent memorized token accuracy perplexity wall-clock time

Baselines: Zero Activations Slimming Hard Concrete Integrated Gradients Greedy SOU Durable Durable-agg Subnet Spectral norm regularization Loss truncation

Builds on: Greedy unlearning (Chang et al.) Durable backdoors Subnetwork discovery via supermasks (edge-popup) Example-tied dropout Loss truncation Spectral norm regularization

Closest papers by meaning

• Mitigating Memorization Where It Happens (NeurIPS 2026 · similarity 0.91)
• Memorization Sinks: Isolating Memorization during LLM Training (ICML 2025 · similarity 0.91)
• Learning What to Forget: Improving LLM Unlearning via Learned Token-Level Importance (NeurIPS 2026 · similarity 0.90)
• Be like a Goldfish, Don't Memorize! Mitigating Memorization in Generative LLMs (NeurIPS 2024 · similarity 0.90)
• Selective Forgetting: Advancing Machine Unlearning Techniques and Evaluation in Language Models (AAAI 2025 · similarity 0.90)
• Constrained Entropic Unlearning: A Primal-Dual Framework for Large Language Models (NeurIPS 2025 · similarity 0.90)