🧮 TinyGSM/GSM8K results

Before, we mainly tested unconditional generation: does the model write good text fast?

Now we also ask: can it reason through math problems?

So we added TinyGSM/GSM8K conditional generation, using the S-FLM eval setup/codebase from @jdeschena.

🔗Blog + paper + code + checkpoints:

📝 Blog: https://david-cripto.github.io/Bio-page/blog/idlm/ 📄 Paper: https://arxiv.org/abs/2602.19066 💻 Code: https://github.com/David-cripto/IDLM 🤗 Checkpoints: https://huggingface.co/kekchpek/models

We updated and released all the materials needed to explore and reproduce IDLM.

⚖️ More baselines + metrics

We expanded the evaluation with recent baselines: ELF(@HuLillian39250, @linluqiu), D-MMD (@emiel_hoogeboom), FLM/FMLM (@chandavidlee , @jw9730) and more. We also added MAUVE and GM metrics.

I’m additionally tagging a few people who might be interested in the new results and materials:

@yjelid @ssahoo_ @dvruette @AndrewC_ML @ValentinDeBort1 @ArnaudDoucet1 @WeiGuo01

📉 GenPPL–entropy frontier.

Motivated by @Sam_Acqua’s recent observation that GenPPL can trade off with entropy under temperature sweeps, we add a GenPPL–entropy plot. Notably, ELF performs very well on this frontier even without using a distillation technique.

(Method / Steps)

📝 New IDLM blog post

We’re also excited to share a new IDLM blog post with intuition behind the method. How IDLM turns many-step DLMs into few-step generators.

@LiDavid2002 @HuLillian39250 @linluqiu @chandavidlee @jw9730 Super cool

@LiDavid2002 @ssahoo_ @dvruette @AndrewC_ML @ValentinDeBort1 @ArnaudDoucet1 @WeiGuo01 Very cool stuff congrats David!

@emiel_hoogeboom @HuLillian39250 @linluqiu @chandavidlee @jw9730 Thanks a lot😃

@yjelid @ssahoo_ @dvruette @AndrewC_ML @ValentinDeBort1 @ArnaudDoucet1 @WeiGuo01 Thanks 😀