Less Loss, Fixed Weights: Logit Corrections for Forgetting
cs.LG, cs.CL
Submitted: 2026-10-30
Updated: 2026-10-30
Comments: 8 pages, 3 figures, 2 tables. Code and data: https://github.com/SeverinVisionary/prior-shift-diagnostic
License: http://creativecommons.org/licenses/by/4.0/
The gist: When a language model is fine-tuned to follow instructions, or further pretrained on a new subject, it often gets worse at predicting standard web text.
Terminology
Abstract
When a language model is fine-tuned to follow instructions, or further pretrained on a new subject, it often gets worse at predicting standard web text. This is known as "forgetting." We ask whether this loss can be undone by simple, fixed adjustments to the model's final token predictions. Across 14 pairs of base and derived models, from 1B to 7B parameters, we fitted a fixed bonus or penalty for each token (a vocabulary bias), alone or together with a temperature. On text held out from correction fitting, the largest top-4K bias recovery estimate is under 10%, with loss conditional on the retained vocabulary. Combining the bias with a temperature recovers up to 36%. Models further pretrained on math lose the most and recover almost nothing. Finally, in a controlled simulation where a model is damaged at random, with no planted bias or temperature change, the bias-plus-temperature fix still recovers 29%. Most of the real-model loss increase stays out of reach of our restricted fits.
Sources
- Large Language Models Are Not Robust Multiple Choice Selectors
- How to Teach Large Multimodal Models New Skills
- Navigating the Alignment-Calibration Trade-off: A Pareto-Superior Frontier via Model Merging
- Learning Dynamics of Continual Learning: A Unified View of Data Attribution, Forgetting, and Plasticity Loss
- Frustratingly Simple Black-Box Adaptation of Language Models via Logit Bias
- Understanding Catastrophic Forgetting in Language Models via Implicit Inference
- When Forgetting is not Catastrophic: On the Mechanics of Spurious Forgetting
Related papers
- A Physics-Regulated Neural Framework for Learning 3D Grain Growth Dynamics
- FlexMoE: Learning One-for-All Multi-Budget Expert Pruning for Flexible MoE Deployment
- OptiMind: Improving Small Language Models for Optimization Modeling through Formulation Guidance and Multi-Turn Learning
- Staying in the Sweet Spot: Boost Reasoning Evolution via Responsive Rollout Difficulty Manipulation
- Examining the Impact of Data and Training Methods on Emergent and Subliminal Misalignment
- Rethinking Black-Box On-Policy Distillation: A Rubric-based Analysis