daniel csillag

Applied mathematician working on machine learning, statistics and compilers.
Currently doing research at FGV EMAp.

[daniel csillag @dccsillag]

@DimitrisPapail (very interesting result! though I do wonder what it would look like under signum rather than SGD)

[daniel csillag @dccsillag]

@DimitrisPapail seems like SGD+LR schedule with restarting, probably even not too hard to analyze theoretically

[daniel csillag @dccsillag]

Oh, and I'll be presenting our paper on Differentially Private E-Values! Poster session 2 (May 3rd), #29

[daniel csillag @dccsillag]

on my way to AISTATS! Happy to talk UQ, e-values, optimization and more. If anyone wants to meet, feel free to shoot me a message

[daniel csillag @dccsillag]

@valentin_kilian Thank you!

[daniel csillag @dccsillag]

Two (out of two) papers accepted at ICML:

- Extending Prediction-Powered Inference through Conformal Prediction
- Avoid What You Know: Divergent Trajectory Balance for GFlowNets

[daniel csillag @dccsillag]

@Gauthier_E_ Thanks, same to you!

[daniel csillag @dccsillag]

Links:

- https://t.co/2A3dwVkaG2 -- work done in collaboration with Pedro Dall'Antonia, Claudio Struchiner and Guilherme T. Goedert
- https://t.co/jAbU1yFXMK -- work done in collaboration with Pedro Dall'Antonia [first author], Tiago da Silva, Salem Lahlou and Diego Mesquita

[daniel csillag @dccsillag]

will be at ICLR! If anyone wants to meet, feel free to shoot me a message

[Aaron Roth @Aaroth]

@lreyzin We just need to start writing better papers. In the new equilibrium, nothing that that GPT can prove directly will be publication worthy. But there will be a larger set of things that we can prove with 6 months worth of effort using gpt.

[daniel csillag @dccsillag]

@Laz4rz agreed. everyone should be using three-space indenting

[daniel csillag @dccsillag]

@kgourg that's because the other rest of eternity was spent writing to stdout due to a missed semicolon

[Francesco Orabona @bremen79]

https://t.co/6nUqJZKJlP

[daniel csillag @dccsillag]

Work done in collaboration with @wkly_infrmtive
Paper link: https://t.co/qjVnWUa0Ps

[daniel csillag @dccsillag]

Now accepted at AISTATS: Differentially Private E-Values!

TL;DR: we introduce mechanisms to convert any e-value into a differentially private version of itself, while preserving its statistical properties

[daniel csillag @dccsillag]

yikes.....

[Alex Cui @alexcdot]

Okay so, we just found that over 50 papers published at @Neurips 2025 have AI hallucinations

I don't think people realize how bad the slop is right now

It's not just that researchers from @GoogleDeepMind, @Meta, @MIT, @Cambridge_Uni are using AI - they allowed LLMs to generate hallucinations in their papers and didn't notice at all.

It's insane that these made it through peer review👇

[daniel csillag @dccsillag]

automatic and silent rank promotion&broadcasting is the root of all evil and should never have been a thing.

just do `export JAX_NUMPY_RANK_PROMOTION=raise` and be free

[daniel csillag @dccsillag]

TIL that JAX has a way to disable/warn on NumPy rank promotion broadcasting!! Actual lifechanger.

[daniel csillag @dccsillag]

@_sanjoydas Regarding novelty, I personally haven't seen such a result before, but to be fair I've seen fairly little theory on optimizing compilers.

[daniel csillag @dccsillag]

@_sanjoydas I think it would be better to be more formal wrt what you mean by programs&compilers here. Due to the 'quine-like' structure of the program P, your definition of a program must be powerful enough to do that. Maybe try working atop some (perhaps augmented) lambda calculus?