Temoor Tanveer

Researcher @ QCRI
CS '24 @ CMU
Working on using AI for discovering algorithms.

[Temoor Tanveer @ttanvali]

What if AlphaEvolve-like systems didn't have to cost a fortune?

LEVI outperforms other frameworks (e.g. ShinkaEvolve, GEPA) on every ADRS problem, at 3-7x less cost, all while using a small QWEN 30B for most steps! In controlled experiments, it reaches the same scores in 10-12x fewer evals.

The trick: invest in the search harness, not the model.

Code: https://t.co/SDolEt247y

Check out the blog below!

[AI-Driven Research for Systems @ai4research_ucb]

🎯 LEVI makes AI-driven algorithm discovery cheaper

[ADRS Blog #19] We feature LEVI, a framework for LLM-based optimization that outperforms prior ADRS frameworks while being ~3–7× cheaper on the main benchmark comparison. Instead of using the biggest models for every step, LEVI uses cheap models for most mutations and reserves frontier models for rare paradigm shifts.

✍️ Read the blog: https://t.co/jExx3t2wwx

📖 ADRS Blog Series: https://t.co/UxujLFWpiD

📄 ADRS Paper: https://t.co/GocF4Xg6yk

👩‍💻 LEVI Code: https://t.co/Zd9RAE18I2

[Temoor Tanveer @ttanvali]

@DimitrisPapail Cool result!

In LEVI (https://t.co/LVaZBllG8G), we show its not just important to recover from basins, but investing in diversity early on is also very powerful.

[Temoor Tanveer @ttanvali]

@OfirPress I think a big part of what linus and others did incredibly well wasn't coding something specific. But rather pulling out a consistent and accurate specification of what needs to built through a lot iteration.

I think this is is exactly what models lack and cannot speedrun.

[Dimitris Papailiopoulos @DimitrisPapail]

One of the worst forms of brainrot AI has cultivated is pessimism about basic research, ie the idea that important work can only happen inside a frontier/neo lab and only with 10k+ GPUs, so the rest should not even bother. What a bleak way to think about science. And it's false.

[Temoor Tanveer @ttanvali]

@maksym_andr But is it more of a miscalibration on the capabilities of these models or our understanding of these problems and their underlying difficulty?

[Temoor Tanveer @ttanvali]

@menhguin claude inducing a coma when the subscription payment fails

[Temoor Tanveer @ttanvali]

I just find this so cool. That different models can be separated this way. Intuitively it makes sense ofc, each has its own quirks

But I wonder how humans map here. I'm sure each one of use has their own em-dashes or "it's not just...". Would be interested in finding out my own.

Not to mention how does the delta between different human-human compare to LLM-LLM? What about humans from one region compared to another?

I think work like this spans beyond just AI writing detection.

[Elyas Masrour @ COLM 2026 // SF @elyasbuilds]

Did you know?

Pangram learns the difference between Claude, ChatGPT, and Gemini in its internal representations, even without being trained on it!

This signal is increasingly recoverable throughout the network, reaching 91% accuracy on a simple linear probe!

[Temoor Tanveer @ttanvali]

@deephivex Cool! I remember the special tokens in the bpe being especially annoying

Should try to get back into tho, was a lot of fun

[Temoor Tanveer @ttanvali]

Very unique perspective on the problem. Interestingly solutions across continual pre training or context management don't seem satisfactory either.

Would be interested in seeing the problem posed as a prompt optimization (e.g. for notes) to see how well GEPA/LEVI do.

[Jacob X. Li @jacobli99]

Continual learning is widely discussed right now, but mostly as improving on the job or avoiding catastrophic forgetting. But it has a different, difficult, and already urgent form:

Given nothing but a corpus of documents, how should AI systems develop expertise in a new, unfamiliar domain? We call this problem Machine Studying.

[Temoor Tanveer @ttanvali]

@_MattJiang_ @UMich You may find the MAS work relevant: https://t.co/Lyhb1ZNrzZ

[Temoor Tanveer @ttanvali]

@prajdabre New teacher (Claude) to student (human) distillation attack incoming

[Temoor Tanveer @ttanvali]

I wonder if those times will ever come back as we move up the abstraction, tools become standardized, and inference becomes faster.

Instead of waiting a few mins each turn and reviewing if the "compiler" did it's job correctly

[Ryan Dahl @rough__sea]

I miss the solitude of programming. Coffee at hand, excitement powering my fingers. Ideally disconnected from the internet.

But I aspire to be an engineer, not a poet. There are new problems and new skills to learn.

[Temoor Tanveer @ttanvali]

@eurofounder Proof LLMs are accurate world models

[Temoor Tanveer @ttanvali]

I meant even less vanilla like OpenEvolve or ShinkaEvolve.

From my experience while working on LEVI, you can even go as small as a QWEN 30B and outperform largers models. Tho the evaluator need to be much tighter and less reward hackable.

If I get a chance tomorrow might try it out with LEVI myself, curious to see if it's possible

[Lorenzo Xiao @ COLM 2026 @lrzneedresearch]

I will be presenting this at July 9th in #ICML2026

TLDR: We believe that Current AI welfare research is deemed to be Frankfurtian Bullshit.

Say Hi to me and @gordondaitianqi

Also: interested in talking about future of work, personalization, privacy, and human centered AI !

[Temoor Tanveer @ttanvali]

@lrzneedresearch @gordondaitianqi Cool title lol

[Temoor Tanveer @ttanvali]

@raveeshbhalla This seems like good news, no?

[Temoor Tanveer @ttanvali]

@morgymcg @iamtrask Yes!

GitHub: https://t.co/18h0KLt5e5
Arxiv: https://t.co/LVaZBllG8G

[Temoor Tanveer @ttanvali]

@novasarc01 There was some work on dynamic benchmarking a couple of years ago, not sure why it didn't catch on more

https://t.co/JpRIsFWdhh

[Temoor Tanveer @ttanvali]

@blelbach Tried this over a couple of systems research problems and noticed a lot less convergence with runs that could last for very long (https://t.co/vzX2NQBSOX).

But obviously you're going in more blind, with a lot less revisited dead ends and learned lessons.

[Temoor Tanveer @ttanvali]

@blelbach What about making the solution archive itself the only memory? MAP-Elites style where every iteration just sees a couple of parents + scores.

From my experience this is a lot less prone to degradation or convergence over longer runs...