AI Paper Atlas

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

TS$^2$: Training with Sparsemax+, Testing with Softmax for Accurate and Diverse LLM Fine-Tuning

XuZiyang, Ananthu Rajendran Pillai, Yinghua Yao, Yuangang Pan

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

This paper tried a common change: Add label smoothing

add label smoothing (0.2) to CE · AlpacaEval BoN@32 · win rate (%): 29.77 -> 28.25 Hurt

What 190 accepted papers report for this change

See this change in detailMatched by the wording of each paper's own result, read by AI from the full text. Papers report gains more often than failures, so this shows what gets published, not the chance a change works.

Read the paper

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 Sparsemax+ and decode with softmax instead of cross-entropyLlama-3.1-8B SFT · AlpacaEval BoN@32win rate (%): 29.77 -> 33.12Improved
train with Sparsemax+ and decode with softmax instead of cross-entropyQwen-2-7B SFT · AlpacaEval BoN@32win rate (%): 31.41 -> 37.48Improved
use TS2 instead of GEMLlama-3.1-8B SFT · AlpacaEval BoN@32win rate (%): 31.53 -> 33.12Improved
use TS2 instead of CELlama-3.1-8B SFT · HumanEval pass@100pass rate (%): 72.60 -> 87.00Improved
use TS2 instead of CEQwen-2-7B SFT · HumanEval pass@100pass rate (%): 86.60 -> 91.50Improved
use TS2 instead of GEMQwen-2-7B SFT · HumanEval pass@1pass rate (%): 41.80 -> 42.20 (CE 44.90)Hurt
use TS2 instead of GEMLlama-3.1-8B SFT · OpenLLM six-task average BoN@32accuracy (%): 75.69 -> 88.88Improved
use TS2 instead of GEM for diversityLlama-3.1-8B SFT · AlpacaEval N-gram diversitydiversity: 20.32 -> 23.78Improved
add suppression term to sparsemax decoupled training (full TS2 vs Decoupling Only) (ablation)TS2 · AlpacaEvalwin rateImproved
decode with softmax instead of sparsemax at inference (vs Unified Sparsemax) (ablation)TS2 · AlpacaEvaldiversityImproved
add sparsemax support set to suppression penalty (vs Suppression Only on CE) (ablation)CE with suppression · AlpacaEvalwin rate and diversityImproved
increase suppression weight alpha from 0.25 to 5.00 (ablation)TS2 · AlpacaEval BoN@32win rate (%): 33.12 -> 25.34Hurt

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