Yash

ML engineer & Independent Researcher (RL, mech interp), fighting CUDA, DMs Open

[Yash @yashetal]

This is really interesting work

[Akshat Gupta @ ACL2026 @akshatgupta57]

🚀New Paper - “Evolutionary Strategies Lead to Catastrophic Forgetting in LLMs"

Continual learning is coming. However, gradient-based algorithms will restrict its use cases. Evolutionary Strategies are resurging as a gradient-free alternative to GRPO… until they forget!

Our work shows that “Evolutionary Strategies Lead to Catastrophic Forgetting in LLMs."

Paper link : https://t.co/P7sB35JNVX

@GopalaSpeech @berkeley_ai

[Yash @yashetal]

System K is intentionally weak, and it does not validate
"Necessities are necessarily necessary"

[Yash @yashetal]

https://t.co/C9BC94HTEd

[Yash @yashetal]

Is it just the inability to deal with emotions, allowing humans to not gain a new perspective from a machine-generated weird (weird here, according to the human, not personally) text/image/video and thats why they will always come up with names like "vibe" coding, "clanker", "ai slop" , but if we look at it from a different dimension being perspective, it will all fit together as the humans did into the picture

[Yash @yashetal]

this is exactly i was talking about the perception of people on AI slop, its nothing its just a human bias to see things and taking it as offensive, its evolutionary mech in humans that benefits survival but i wonder whether is this the thing that will be the biggest bottleneck while we are on a path to achieve AGI, i think its going to be very difficult to evaluate US rather than AI as we progress

[投机实验室 @LabSpeculation]

最近，有个博主把莫奈的真迹，标记成了AI画作，让网友尽可能详细地描述，这张图到底差在哪里。

结果这幅画被网友狂喷。
甚至有网友直言这幅画是SHIT，各路人马疯狂贬低，从各个专业角度锐评AI创作拙劣至极。

当博主公布答案这是真的是莫奈的画之后，很多人默默删除了评论。

现在只要一张图被打上 AI 标签，很多人在认真看之前，结论就已经先出现了。

[Yash @yashetal]

Reading math/computational logic LO paper feels so historic and ancient like magic spells books (in a good way) and meanwhile reading these new Agentic/LLM paper feels modern utopian projection onto future
Both have logic at its core but feels like two different worlds

[Machine Learning (ML) Papers @Memoirs]

Learning dynamical systems from noisy data with Weak-form Kernel Ridge Regression

Max Kreider, John Harlim, Daning Huang
https://t.co/hdFG2OWJ4M [𝚌𝚜.𝙻𝙶 𝚖𝚊𝚝𝚑.𝙳𝚂]

[Yash @yashetal]

coding the whole solution for a complex problem and then explaining the thought process on how you arrived at the solution is much more efficient in the interview process and less time consuming + better experience rather than explaining the thoughts as they come while coding up the solution cause the thoughts are messy (and its even more non-simplistic in iq above 125 ) so it might sound very sketchy to the one who is more proned to the simplistic dwell down logic solutions, also they will end up hiring for the folk who can do simple tasks faster but cannot do complex tasks faster

[Yash @yashetal]

This is a good trend

[BioTender @Junioryu136689]

I curated 315 AI × Bio / Biomedical papers from ICML 2026 and open-sourced the full collection on GitHub.

🧬 315 papers
⭐ 27 Spotlights
💻 161 papers with code
📚 10 research categories
From protein design and drug discovery to genomics, single-cell biology, neuroscience, and clinical AI.
A searchable map of AI × Bio at ICML 2026.
⭐ Stars are appreciated!
GitHub:
https://t.co/W3tcdEPDrW

[Yash @yashetal]

@gabriberton Yes I don't know why people are falling for this, maybe it worked in some cases but for me I tried it 3-4 times and it gave me not expected results

[Yash @yashetal]

It's weird to think about though that we need prompts still to steer the model in a particular direction we want instead of the model intelligent guessing

[Rohan Paul @rohanpaul_ai]

This is the prompt Yale and Univ of Chicago researchers used when asking LLMs for new research ideas.

Feed LLMs prior work, ask for ideas, then measure how repetitive the ideas get.

The surprising finding is that LLMs often treat research ideation as connecting what already exists, while humans use a wider set of problem-finding moves.

LLM-generated ideas reveal a bias toward safe bridge-and-combine proposals.

[Yash @yashetal]

If you are doing certification in 2026 it's useless completely tbh

[Phil Chen @philhchen]

https://t.co/oMoEAYkqGl

[Yash @yashetal]

This is useful stuff

[JFPuget 🇫🇷🇺🇦🇨🇦🇬🇱 @JFPuget]

My team open sourced a set of agent skills to run an entire @kaggle competition workflow from plain language. We ship an easy to install plugin for your favorite coding agent.

👉 Try it now: https://t.co/ayVnzYycLY

[Yash @yashetal]

Big picture: Moves beyond retrieval or post-hoc graphs → makes the model's native reasoning symbolic, relational, and reusable. Potential for scientific agents that build and iterate on shared knowledge graphs.

[Yash @yashetal]

Just dropped on arXiv: "Graph-Native Reinforcement Learning Enables Traceable Scientific Hypothesis Generation through Conceptual Recombination."

A new family of models (Graph-PRefLexOR) that turns LLMs into structured, inspectable hypothesis generators for materials science & mechanics.

Link: https://t.co/ppuDFxh7aq

[Yash @yashetal]

Results on 100 challenging open-ended questions from materials/mechanics literature:

-> 40-65% gains over base models

-> Biggest wins in reasoning traceability (makes the thought process inspectable & reusable)

-> ~2-3× greater semantic diversity

-> Better alignment between intermediate graphs and final answers

Cool analyses:

-> Embedding trajectories show more organized, directional exploration

-> Test-time graph expansion: extra compute → long-range conceptual recombination (not just broader search) within a bounded semantic space

Why this matters for materials design:

Properties emerge from multi-scale, cross-domain mechanisms (molecules → mesoscale → processing → performance). Graph-native reasoning forces the model to explicitly connect these instead of hand-waving in linear text.