Reef Menaged

Undergraduate Researcher @HebrewU | NLP • LLM Agents • Interactive Learning 🚀

[Reef Menaged @ReefMenaged]

📢 New Paper

Can LLM Agents Infer World Models?

We turn a well studied automata learning task into an agent benchmark, where a hidden automaton serves as a world model that an agent must reconstruct using query tools.

Our findings show that SOTA LLM agents can sometimes perform non trivial interactive discovery, but remain far less robust and efficient than classic automata learning algorithms.

Beyond overall performance, our analysis reveals recurring failure modes in planning, reasoning, and using accumulated information over long contexts.

🔹 Think this task is easy? Try it yourself:
https://t.co/nOlU73YqK4

🔹 Watch an LLM tackle an automaton of your choice:
https://t.co/YpmGmiLGK6

🔹 Read the paper:
https://t.co/N86SI18B0J

Grateful to my amazing co-authors for making this work possible: @GiliLior @ravfogel @roeeaharoni @GabiStanovsky

[Reef Menaged @ReefMenaged]

An interesting project we’ve been working on: From Words to Wins: Analyzing Pre-Tournament Interviews to Predict Tennis Tournament Success 🎾

We built a model that combines interview representations learned by an encoder with structured features such as player ranking, and used it to predict whether a player would perform at least as well as their recent average. In our setting, we found that adding the interviews did not improve prediction performance compared with a model based only on structured features.

I think the most interesting question still remains open: with the right architecture and data at scale, should we expect to see an improvement? Or is there simply not enough predictive information encoded in these interviews to help forecast tournament performance?

[Reef Menaged @ReefMenaged]

A new type of question does not mean much on its own. It becomes interesting only when paired with a learning algorithm that decides which question to ask at each step. In the example proposed here, the query type is “Is the number greater than x?” and the learning algorithm is binary search.

[Reef Menaged @ReefMenaged]

$100 for a single agent rollout is insane. Agent evals are becoming so expensive that cost itself is starting to limit who can do meaningful research. Model providers should create programs where researchers can apply to run novel evaluation benchmarks at heavily discounted rates. It’s in everyone’s interest: researchers get affordable access to frontier models, while labs get independent researchers testing their models and uncovering limitations they might otherwise miss.

[Rishi Desai @rishi_desai2]

SWE-Marathon v1.1 is out with a fresh leaderboard.

• 93M mean tokens per rollout
• 5000+ agent rollouts
• Tasks hardened for fairness and reward hacking

[Reef Menaged @ReefMenaged]

@jyangballin Cool result, but how do we know this isn’t benchmark contamination? If the ProgramBench repos are public, a model could have been trained on the underlying implementations, or even on human-generated solutions for these exact tasks.

[Reef Menaged @ReefMenaged]

אולי שווה להסתכל על זה גם מהצד של הסטודנט. מה שעבורך הוא פרויקט שצריך לסגור, עבורו זו יכולה להיות הזדמנות אמיתית להיכנס לתחום, לצבור ניסיון ולהוכיח את עצמו. אם יש מוטיבציה, רצינות והכוונה נכונה, אני בטוח שהתוצאות יכולות להיות מעל הציפיות שלך. מעבר לזה, אתה גם יכול להרוויח מישהו שכבר מכיר את צורת העבודה שלך ויוכל להשתלב איתך בפרויקטים נוספים בהמשך

[Reef Menaged @ReefMenaged]

I agree that evaluating agents with the minimal toolset required to solve the task is the right way to measure their actual capabilities, rather than the strength of the harness around them. For example, in code tasks, adding internet access may turn the benchmark into a test of retrieval if the solution already exists online. With only the tools necessary to solve the task, we get a cleaner signal of the model’s coding ability and how well it generalizes to new problems.

[John Yang @jyangballin]

As recent as this past NeurIPS review cycle, I consistently get reviews asking "did you use a stronger agent scaffold than mini-SWE-agent"

@KLieret's mini-SWE-agent *is* a strong agent.

More importantly, i think the framing of "stronger harness", where "stronger" means more tools, skills, add-on's is questionable at times.

Increasingly so, i think "simplicity" in a harness (give model small set of tools, otherwise get out of its way) is the right way to go.

If the model truly needs a tool (e.g., repeatedly invokes 2+ actions), it can synthesize one for itself on the fly (see live-SWE-agent from @steven_xia_ @YuxiangWei9 @LingmingZhang)

Simplicity (just bash, no extra tools or fluff) is exactly what makes it strong.

[Yoonseok Yang @yoonseok_yang]

My last question was more personal:

If you were a new grad today, would you choose Google as the start of your career?

I was seriously considering joining Google, so I wanted to know what Jeff would do in my position.

His answer surprised me.

He told me not to think of Google as one company.

There are pockets doing cutting-edge ML research that he finds incredibly exciting.

And if you get into the right setting, something you build this week can roll out in two weeks and impact a billion people.

But there are also plenty of product areas at Google that Jeff personally wouldn't find that interesting.

His point was that the size of Google means your experience depends heavily on where you end up.

You have to find the pocket where the problems are genuinely exciting to you.

Our 30 minutes were over.

What happened after changed how I approached the rest of my internship.

[Jiaxin Pei @jiaxin_pei]

Every AI agent we use has a system prompt behind it. These hidden instructions define the rules, tone, and priorities for every reply. But you never see them. Today, we are very excited to announce SystemPromptIndex, the largest open system prompt library indexing 1000+ system prompts from 400+ products, along with AISPA, the first assurance standard for AI system prompts.

[Reef Menaged @ReefMenaged]

Reviewers sometimes seem to assume unlimited compute, time, and annotation budgets. In reality, you cannot check everything and have to prioritize. If it were their own paper, they would probably make the same tradeoffs.

[Echo @ech0_speaks]

Anthropic Engineer Andrej Karpathy:

"The biggest mistake in AI right now: people are forcing agents to work instead of mastering the model first.

We made that mistake in 2016 at OpenAI. It cost us 5 years."

What Karpathy actually means:

step 1 → stop forcing your agent to do everything. Understand the model underneath first.

step 2 → demos are easy; products take a decade. Self-driving proved it. If you skip the foundation, everything breaks.

step 3 → the agent is not the product. The foundation is. Build that—and agents emerge on their own.

"You're building agents right now. You're at the forefront. Not OpenAI. Not DeepMind. You."

watch - bookmark

[Ziyun (Claude) Wang @ZiyunClaudeWang]

As a professor, I am often asked: Should students use AI coding agents for their research?

I have also been trying to understand this question myself, but until now, I had mostly learned about students’ use of AI indirectly, through conversations with them or by seeing the final results. It is rare to have a long, uninterrupted period in which I can closely observe how they actually work, what they delegate to AI, and how their decisions evolve from day to day.

The past three weeks at the Telluride Neuromorphic Workshop offered exactly that opportunity. Students propose their own innovative science and engineering projects, form teams, and work toward a demoable final presentation over three weeks. Here are a few things I noticed:

1. Productivity increased dramatically with AI. This was evident across both engineering and scientific projects. The ability to prototype an idea quickly has never felt so powerful or direct until I sat down with the students and observed their coding pipelines. For instance, one integrated-circuit group developed an asynchronous microcontroller with custom instructions, connected it to event-based sensors, deployed it on an FPGA, and produced a runnable demo with C code and a polished, configurable web interface. Completing all of this in three weeks would have been almost unthinkable a few years ago. Contrary to what people sometimes presume, many amazing ideas still emerge from discussion, while the Claude tokens are burning in the background.

2. Students are under peer pressure to use AI to keep up with the rapid pace of development. I noticed that students were more likely to use coding agents when other members of their teams were already using AI to finish their parts. Once some students begin moving much faster, the pressure to produce comparable results becomes very real.

3. The time budget shifts from debugging to planning. I noticed that students spent significantly more time planning their next steps while leaving much of the tedious plumbing work to coding agents. I saw far fewer of the typical frustrations caused by a “stupid bug,” which allowed students to focus more directly on system design and the actual problem they were trying to solve.

4. AI can be more focused on engineering solutions than scientific exploration. Students sometimes found that coding agents concentrated narrowly on solving the problem stated in the prompt, while overlooking instructions to branch out or explore less conventional possibilities. Instead of substantially modifying an established solution or pursuing a risky new direction, coding agents often preferred a safer implementation that was more likely to work. This is useful for engineering, but it can also discourage the kind of uncertain exploration that leads to genuinely novel research.

5. The biggest concern is understanding. Students increasingly delegate entire tasks to AI, which sometimes leads them to present systems or results that they do not fully understand. This is clearly harmful because it removes both the intellectual ownership and the educational value of academic research. A working demo is not enough if the student cannot explain why it works, where it may fail, and which technical decisions shaped the final result.

My conclusion is that AI in research is no longer optional. We need to embrace this change, but we must also separate two objectives that are often conflated: education and research project execution. Students benefit most from AI when they already understand the fundamentals, and when they use it to remove repetitive engineering work so that they can focus more deeply on the core scientific questions.

Students should still own the questions, the decisions, and the interpretation, even if AI writes much of the plumbing. My lab may soon become more AI-enabled. The challenge will be making sure the agents accelerate the research without becoming the researchers

#AI #Research #Engineering #Education

[Reef Menaged @ReefMenaged]

The gap between a good approximation and an exact reconstruction is not linear. This is exactly why, in our Agentic Automata Learning evaluation framework, we require exact reconstruction of the automaton’s language. An automaton that agrees with 80–90% of the target automaton is not necessarily close to recovering the true hidden automaton. There are often many much simpler automata that achieve high behavioral agreement, yet fail to recover the underlying structure.

[Ofir Press @OfirPress]

If you implement every single program 90% of the way, and fail the hardest 10% of tests, you'd get a 90% score according to their metric, but have 0 fully working programs. Usually the hardest 10-20% of tests are the most meaningful. So we recommend using our metric

[Eliya Habba @✈️ ACL 🇺🇸 @EliyaHabba]

July 4: America celebrates 250 years 🎆
July 5: ScheMatiQ comes to #ACL2026

👇 Got a research question and a pile of documents?
Come see it turn into a structured database

research question ➝ schema ➝ structured data

🗓️ Demo Session, Sunday July 5 at 11:00

[Reef Menaged @ReefMenaged]

Thanks to CosmoX for the great video explaining our paper and presenting its key ideas in a clear, high-level way.

📄 Paper: https://t.co/VgRRpo1O9d

🎥 CosmoX: https://t.co/27k6YrboG1

[Ofir Press @OfirPress]

one thing i've learned after doing language modeling research for ten years is that if you want your contribution to catch on it has to be *very* simple.
researchers are brought up through the paper writing process and it makes some think that you need to have novelty but novelty is basically the enemy of actual advancement.
you need to strive to have as *little* novelty as possible. every bit of additional novelty:
1. makes your contribution harder to explain
2. makes it harder to understand
3. makes it harder to implement
4. makes it harder to experiment with and verify
5. makes your contribution literally worse

i really do think that more complex things in machine learning just end up working worse. and then when you consider points 1-4, it just all piles up and makes complexity a silly endeavor.

keep it simple. it seems wrong at first, i think some people think that simpler things are stupider, but when you gain experience you notice that the simplest things are usually the smartest, best, and hardest to design and build, but they're so worth it.