REVIEW 6 cited by
Understanding Catastrophic Forgetting in Language Models via Implicit Inference
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
read the original abstract
We lack a systematic understanding of the effects of fine-tuning (via methods such as instruction-tuning or reinforcement learning from human feedback), particularly on tasks outside the narrow fine-tuning distribution. In a simplified scenario, we demonstrate that improving performance on tasks within the fine-tuning data distribution comes at the expense of capabilities on other tasks. We hypothesize that language models implicitly infer the task of the prompt and that fine-tuning skews this inference towards tasks in the fine-tuning distribution. To test this, we propose Conjugate Prompting, which artificially makes the task look farther from the fine-tuning distribution while requiring the same capability, and we find that this recovers some of the pretraining capabilities in our synthetic setup. Since real-world fine-tuning distributions are predominantly English, we apply conjugate prompting to recover pretrained capabilities in LLMs by simply translating the prompts to different languages. This allows us to recover in-context learning abilities lost via instruction tuning, natural reasoning capability lost during code fine-tuning, and, more concerningly, harmful content generation suppressed by safety fine-tuning in chatbots like ChatGPT.
Forward citations
Cited by 6 Pith papers
-
Helpfulness Hurts: Domain-Dependent Degradation of Mid-Trained Compassion Values Under Post-Training
Helpfulness SFT/GRPO degrades mid-trained animal compassion far more than coding SFT/GRPO on ANIMA, with the compassion effect transferring cross-lingually while general moral-reasoning gains do not.
-
Transformers Don't In-Context Learn Least Squares Regression
In-context regression transformers do not approximate OLS: they underperform it even in-distribution, fail on out-of-subspace prompts, and their failures correlate with a low-rank spectral signature in the residual stream.
-
Reflect, Retry, Reward: Self-Improving LLMs via Reinforcement Learning
A GRPO-based method that rewards only self-reflection tokens, not answer tokens, improves LLM accuracy on function calling and Countdown math tasks using only binary success/failure feedback.
-
Scaling Point-in-Time Language Models
Scaling point-in-time LLMs to 4B parameters and 1T temporally filtered tokens narrows the gap to unrestricted models to about 8–11 average points and yields positive out-of-sample Sharpe ratios from news embeddings.
-
SOS-LoRA: Static Orthogonal-Subspace Low-Rank Adaptation with Fixed Multi-Scale Scaling
A LoRA update split into several fixed, differently-scaled low-rank experts with orthogonal input directions improves fine-tuning accuracy at the same parameter count.
-
Retaining by Doing: The Role of On-Policy Data in Mitigating Forgetting
RL post-training forgets less than SFT because it trains on on-policy data; refreshing SFT data each epoch also reduces forgetting.
Discussion (0). Sign in to comment.