Aaron Imani

PhD Student in Software Engineering at UC Irvine

[Chris Wendler @wendlerch]

Data is plenty, knowledge is scarce. We began to close this gap thanks to deep learning <3 Neural networks can learn “programs” that often achieve superhuman performance from data alone. What insights are encoded in their weights? Here we took a first step on AI protein folding.

[Neil Zeghidour @neilzegh]

Me defending my O(n^3) solution to the coding interviewer.

[Mustafa @oprydai]

strong men creates C language.
C creates goodtimes.
goodtimes creates python, python creates ai, ai creates vibe coding, vibe coding creates weak men, weak men creates bad times, bad times creates strong men

[Aaron Imani @aaron_imn]

@rohanpaul_ai Thanks for resharing our work, Rohan!

[Rohan Paul @rohanpaul_ai]

This paper shows that LLMs store code comments as a separate idea that can help or hurt.

Performance shifts range from 90% worse to 67% better when that comment idea is pushed down or up.

Comments are text in code that the computer ignores, and the authors test whether an LLM keeps a clear comment footprint inside its layers.

They use Concept Activation Vectors (CAV), which find a direction in the model's hidden number lists that means the code has comments.

They learn it by training a detector to tell commented Java code from the same code with comments removed.

Then they slightly shift those number lists in that direction or the opposite direction, so the model acts like comments exist or not without changing the input text.

Javadoc is the most distinct comment type, and effects vary by task, translation can improve when comment concepts are reduced, but completion and refinement often get worse, and summarization activates comment concepts most while completion activates them least.

----

Paper Link – arxiv. org/abs/2512.16790v1

Paper Title: "Inside Out: Uncovering How Comment Internalization Steers LLMs for Better or Worse"

[Aaron Imani @aaron_imn]

🎉 Our paper, "Inside Out: Uncovering How Comment Internalization Steers LLMs for Better or Worse", has been directly accepted in the second cycle of #ICSE2026. A huge thanks to my amazing co-authors
@iftekhar_ahmedi
and
@MMoshirpour. Stay tuned for the link to our paper!

[Aaron Imani @aaron_imn]

🎉 Our paper, "Context Conquers Parameters: Outperforming Proprietary LLM in Commit Message Generation" has been directly accepted in the second cycle of #ICSE2025. A huge thanks to my amazing co-authors @iftekhar_ahmedi and @MMoshirpour.

arXiv: https://t.co/DbcjWy8n6M

[Aaron Imani @aaron_imn]

I had a great time presenting our paper, “Context Conquers Parameters: Outperforming Proprietary LLMs in Commit Message Generation”, at #ICSE2025!

You can watch the pre-recorded presentation here:
https://t.co/wBodxgMc74

#SoftwareEngineering

[Aaron Imani @aaron_imn]

🎉 Our paper, "Context Conquers Parameters: Outperforming Proprietary LLM in Commit Message Generation" has been directly accepted in the second cycle of #ICSE2025. A huge thanks to my amazing co-authors @iftekhar_ahmedi and @MMoshirpour.

arXiv: https://t.co/DbcjWy8n6M

[Mahan Tafreshipour @mahan_tp]

How do developers manage and evolve prompts in LLM-integrated software?🤔✍️

Our #MSR2025 paper (w/ @aaron_imn @EricHuang4312 @esaufba @tomzimmermann @iftekhar_ahmedi) reveals key insights from analyzing prompt changes across 243 open-source repos!🔍✨

https://t.co/haOjznkhAc

[Aaron Imani @aaron_imn]

@jaydeb_sarker @iftekhar_ahmedi @MMoshirpour Thanks, Jaydeb!

[Aaron Imani @aaron_imn]

🎉 Our paper, "Context Conquers Parameters: Outperforming Proprietary LLM in Commit Message Generation" has been directly accepted in the second cycle of #ICSE2025. A huge thanks to my amazing co-authors @iftekhar_ahmedi and @MMoshirpour.

arXiv: https://t.co/DbcjWy8n6M