AI Paper Atlas

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

Enhancing Trustworthiness of Fine-Tuned LLMs via Regularized Subset Selection

Kumar Shubham, Nishant Sharma, Karn Tiwari, Prathosh AP

ICLR 2026 · Large language models › Instruction and fine-tuning · approach: Efficient LLM fine-tuning systems

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
apply PBRF repair on selected detrimental subset after SFTQwen2.5-7B post-SFT · Truthfulnesslog-odds: 0.732 -> 0.662Improved
apply PBRF repair on selected detrimental subset after SFTPythia-6.9B post-SFT · Stereotypical biaslog-odds: −0.380 -> −0.449Improved
apply PBRF repair after SFTQwen2.5-7B post-SFT · Machine ethicslog-odds: −0.253 -> −0.275Improved
apply PBRF repair after SFTpost-SFT models · Anthropic HH perplexityperplexity: 5.401 -> 5.419No clear effect
repair with a common subset for all trust metrics instead of metric-specific subsetsSeerRepair common subset · Truthfulnessrelative log-odds change %: +21.73Mixed
add PBRF-based update instead of GA/SGA/GA+KL gradient ascent (ablation)trust repair · trust metricslog-odds and perplexityImproved
select subset with DPP/attribution instead of random subset (ablation)gradient ascent repair · Machine ethics, Pythia-6.9Bperplexity relative change %: -26.2Improved
add DPP diversity regularization to subset selection (ablation)repair method, Pythia-1.4B and Qwen2.5-1.5B · higher learning ratesperplexityImproved
use PBRF repair instead of retraining without detrimental samplesPythia-1.4B · Truthfulnesslog-odds: 0.504 -> 0.476No clear effect
use repair instead of DPOQwen2.5-7B · Truthfulnesslog-odds: 0.787 -> 0.573Improved
use repair instead of BiPO steeringPythia-1.4B · Stereotypical biasrelative log-odds change %: 4.339 -> 13.430Improved
apply repair to Qwen2.5 modelsQwen2.5-7B · GSM8K accuracyaccuracy: 80.89 -> 80.14No clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy