AI Paper Atlas

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

Supervised Fine-Tuning or Contrastive Learning? Towards Better Multimodal LLM Reranking

Xin Zhang, Ziqi Dai, Mingxin Li, Yanzhao Zhang, Dingkun Long, Pengjun Xie, Meishan Zhang, Wenjie Li, Min Zhang

ICLR 2026 · Large language models › Instruction and fine-tuning · approach: Fine-tuning generalization and delta compression · 6 citations (OpenAlex)

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
train with SFT yes/no classification instead of contrastive InfoNCE (ablation)MLLM reranker · MRBaverage score: 56.40 -> 58.09Improved
use SFT weight instead of CL weight (direction fixed) (ablation)MLLM reranker · MRBaverage score: 56.99 -> 58.09 (D_SFT); 56.40 -> 57.88 (D_CL)Improved
use SFT direction instead of CL direction (weight fixed) (ablation)MLLM reranker · MRBaverage score: 57.88 -> 58.09 (W_SFT); 56.40 -> 56.99 (W_CL)Improved
use constant weight 1 instead of instance-specific weights (ablation)SFT reranker · MRBaverage score: 58.19 -> 49.47Hurt
add tau mask that stops learning on well-learned pairs (ablation)constant-weight baseline · MRBaverage score: 49.47 -> 56.57Improved
replace SFT direction with random-initialized learnable projection (ablation)SFT reranker · MRBaverage score: 58.09 -> 56.75Hurt
use random-initialized learnable projection as CL direction (ablation)CL reranker · MRBaverage score: 56.40 -> 57.72Improved
expand SFT target tokens from yes/no to top-16 tokens (ablation)SFT reranker · MRBaverage scoreNo clear effect
compute CL loss in FP32 instead of FP16 (ablation)CL reranker · MRBaverage score: 56.09 -> 56.40Improved
freeze LM head (ablation)SFT reranker · MRBaverage scoreNo clear effect
increase number of negatives per sample up to 16 (ablation)SFT reranker · MRBaverage scoreImproved
rerank with GMR-7B instead of GME-2B retrieval onlyGME-2B top-100 · MRBaverage score: 52.54 -> 63.85Improved

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