When a large language model is fine-tuned to specialize in an area like finance, it often loses general skills like reasoning or following instructions. The researchers compared ways of fine-tuning with low-rank adaptation (LoRA): training on correct answers only, on reasoning steps from a larger model, or on separate adapters for steps and answers. They also compressed those steps. Training on answers alone improved the finance task the most while keeping general capabilities near the original. Adding reasoning steps did not help more. Compressing steps partly restored calibration but not fully. The lesson: simpler fine-tuning can be better for keeping a model versatile.
Language models can learn financial labels better without teacher explanations
In a financial classification study, training on labels alone outperformed training with teacher-generated reasoning while preserving general capabilities near the original model’s level.
Big Tech
Saleh Ashkboos · Dhruv Desai · Lavinia Ghita
NVIDIA
Research Digest··2 min read
Thread:Alignment Training Dynamics
In one line
Response-only LoRA fine-tuning improves domain performance while preserving general capabilities better than incorporating reasoning traces.
Before you read this
Two ideas this paper builds onconfidence calibrationConfidence calibration aligns a model's predicted confidence with its actual accuracy across predictions.90 papers →context compactionCompressing or summarizing an LLM's past context into a smaller representation to extend memory and improve reasoning within limited token budgets.80 papers →
Read it as
Ashkboos and colleagues compared three ways to adapt Qwen3-8B to financial classification using 200,000 training examples.
The authors used low-rank adaptation, or LoRA, which freezes the original model’s weights and trains small parameter updates called adapters.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.