Arda Uzunoğlu

cs @JohnsHopkins'27
#FreeImamoglu

[Arda Uzunoğlu @aardauzunoglu]

As LLMs surpass humans on many fronts, how can we keep training stronger models?

Our ICML 2026 paper studies this via weak-to-strong generalization and shows that learning when to trust the weak teacher may be key.

Trust Functions: Near-Lossless Weak-to-Strong Generalization 🧵

[Anthropic @AnthropicAI]

Our internal data shows Claude is accelerating AI development—a possible path to recursive self-improvement, or AI autonomously building a more capable successor.

It’s happening faster than we thought, and the implications deserve greater attention. https://t.co/OVVPJO7VQx

[Michael Levin @drmichaellevin]

New #preprint - @YanboZhang3

"Intelligence from Learnable Novelty"
https://t.co/PQz0lPckcL

What if we optimize Epiplexity (https://t.co/YYjWdlqKzn @m_finzi @andrewgwils ) instead of measuring it? We have derived a closed-form approximation of Epiplexity and discovered a deep connection between it and intelligence. This allows us to reinterpret Epiplexity as a form of learnable novelty, providing a brand-new understanding of what intelligence is. By maximizing Epiplexity across various systems, all of them exhibited interesting behaviors:

Cellular Automata: Maximizing Epiplexity directly generates complex soliton interactions similar to Rule 110.
Image Encoders: It automatically causes the encoding to cluster, successfully categorizing different handwritten digits without supervision.
Reinforcement Learning: Introducing Epiplexity improves the performance of PPO in sparse reward tasks.

We also explored the relationship between the theory of learnable novelty, the free energy principle, and novelty search. We hope this work helps us better understand the nature of intelligence and its origins.

[tender @tenderizzation]

it’s 53.411 vs 53.49

[Alvin Zhang @ZAlvin39105]

1/7 Releasing my first blog post, written over the past few days in spare moments at ICML.

I’ve long been interested in test-time training and looped models, and finally had time to play with their intersection:

“Loop deeper, or adapt?”

https://t.co/SQfiOTjGuF

[Marc Marone @ruyimarone]

Check out this poster on weak to strong generalization from very talented undergrad @aardauzunoglu and colleagues from @jhuclsp !

HALL A #4214, today at 10:30!

[Daniel Khashabi 🕊️ @DanielKhashabi]

Imagine this: It's Friday. You've been deep in a debugging session with your agent for hours. Suddenly it says: "Need to compact context" because it's about to exceed the context window. And with that, it forgets various details about your session.

𝐓𝐡𝐞 𝐜𝐡𝐚𝐥𝐥𝐞𝐧𝐠𝐞

As they work, agents' context fills with both useful things and junk (dead-end searches, stale results, superseded plans). That junk impacts everything that follows. Compaction (summarizing the context and resuming from the summary) is the standard fix.

Two questions matter: (1) 𝑤ℎ𝑒𝑛 to compact, and (2) ℎ𝑜𝑤 to compact.

Both matter: a poorly-timed call discards the partial result you still needed.

𝐖𝐡𝐚𝐭 𝐝𝐨 𝐩𝐞𝐨𝐩𝐥𝐞 𝐝𝐨 𝐭𝐨𝐝𝐚𝐲?

Within academia there are a few camps:

1️⃣ Deployed agents compact at a certain token threshold. This is an intuitive starting point: compact if you're about to exceed. But it starts to make less sense once one realizes that this compaction point may appear in the middle of an important task. And there is no agreement on the threshold — triggers range from 30% to 95% of the context window.

2️⃣ Another camp compacts on a fixed interval — every k tokens regardless of trajectory state. This increases the frequency of the compactions, which allows one to use shorter active context (and hence less memory). But again there is no informed logic to when compaction is applied.

3️⃣ Training-based approaches bake the when-to-compact decision into weights via SFT/RL. But with the increasing capabilities of the models, especially with their tool-use capabilities, is training really a necessity?

In industry, various agents have compaction strategies that more or less work. We don't always know how they work. Even in cases that we do, it's unclear how well they do and why.

𝐎𝐮𝐫 𝐩𝐫𝐨𝐩𝐨𝐬𝐚𝐥: 𝐒ᴇʟꜰ𝐂ᴏᴍᴘᴀᴄᴛ

A training-free scaffold pairing two inference-time elements:

(i) 𝐀 𝐜𝐨𝐦𝐩𝐚𝐜𝐭𝐢𝐨𝐧 𝐭𝐨𝐨𝐥. The model emits a summarize signal, the scaffold runs the summarization, and generation resumes from the summarized prefix. This is the ℎ𝑜𝑤.

(ii) 𝐀 𝐫𝐮𝐛𝐫𝐢𝐜 𝐠𝐚𝐭𝐞. At periodic probes, the model returns COMPRESS or CONTINUE. Fire when a sub-task has resolved or the trajectory is converging. Suppress mid-derivation or when stuck. This is the 𝑤ℎ𝑒𝑛.

The same model is generator, judge, and summarizer.

𝐎𝐮𝐫 𝐞𝐯𝐚𝐥𝐮𝐚𝐭𝐢𝐨𝐧𝐬

We evaluate on two task families and 6 datasets: (1) Competition math (IMO-Answerbench, HMMT Nov 25, HMMT Feb 26) and (2) Agentic search (BrowseComp, BrowseComp-Plus, DeepSearchQA).

We also evaluated 7 models: GLM-4.7-Flash, MiniMax-M2.5, MiMo-V2-Flash, Qwen3-4B-Instruct, Qwen3-30B-A3B-Instruct, Qwen3.5-4B, Qwen3.5-9B.

𝐅𝐢𝐧𝐝𝐢𝐧𝐠𝐬

🔹 The pattern of gains is consistent across various evaluations: baseline < fixed-interval ≤ 𝐒ᴇʟꜰ𝐂ᴏᴍᴘᴀᴄᴛ.

🔹 In scenarios where window size is limited, +𝟏𝟖.𝟏 𝐩𝐨𝐢𝐧𝐭𝐬 over no-compaction.

🔹 Even if there is no limit on context window, 𝐒ᴇʟꜰ𝐂ᴏᴍᴘᴀᴄᴛ is still helpful. It is the strongest method on all three models: +𝟓–𝟗 𝐩𝐨𝐢𝐧𝐭𝐬 over no-compaction, at 𝟑𝟎–𝟕𝟎% 𝐥𝐨𝐰𝐞𝐫 𝐩𝐞𝐫-𝐪𝐮𝐞𝐬𝐭𝐢𝐨𝐧 𝐜𝐨𝐬𝐭.

🔹 In our analysis, we find that 𝐒ᴇʟꜰ𝐂ᴏᴍᴘᴀᴄᴛ fires earlier than the fixed-threshold baseline.

🔹 We also observed that 𝐒ᴇʟꜰ𝐂ᴏᴍᴘᴀᴄᴛ helps most on the hardest questions, where deep search accumulates the largest contexts. On easy questions, all three policies are within noise.

——

💡 𝐓𝐚𝐤𝐞𝐚𝐰𝐚𝐲:

💡 "When to compact" is a meta-cognitive capability a lightweight rubric can supply at inference time .

💡 Across 6 benchmarks and 7 models: 𝐒ᴇʟꜰ𝐂ᴏᴍᴘᴀᴄᴛ matches or exceeds fixed-interval summarization at 30–70% lower cost.

💡 To our knowledge, this is the first work to introduce and evaluate rubric-based adaptive compaction timing in a training-free setting.

[Tianjian Li @tli104]

🚨 New paper: "Self-Compacting Language Model Agents"

LM agents build up long traces of reasoning and tool calls. As the trace grows, old mistakes and stale info stick around and anchor everything that follows. We ask: can the model itself decide when to clean up?

[Daniel Khashabi 🕊️ @DanielKhashabi]

Interesting work by @aardauzunoglu and @ZAlvin39105 that frames weak-to-strong generalization as a data selection problem, i..e "which weak labels should the student trust?" (to appear in ICML)

[Alvin Zhang @ZAlvin39105]

https://t.co/Sm46xPE67c

[Arda Uzunoğlu @aardauzunoglu]

This is a joint work by @aardauzunoglu @ZAlvin39105, and supervised by @DanielKhashabi. We also greatly appreciate Tianjian Li and Jack Zhang, as well as the JHU DSAI community for their valuable discussion and feedback!

Paper: https://t.co/4Rkdxtr5gV

[Arda Uzunoğlu @aardauzunoglu]

As LLMs surpass humans on many fronts, how can we keep training stronger models?

Our ICML 2026 paper studies this via weak-to-strong generalization and shows that learning when to trust the weak teacher may be key.

Trust Functions: Near-Lossless Weak-to-Strong Generalization 🧵

[Anthropic @AnthropicAI]

Our internal data shows Claude is accelerating AI development—a possible path to recursive self-improvement, or AI autonomously building a more capable successor.

It’s happening faster than we thought, and the implications deserve greater attention. https://t.co/OVVPJO7VQx

[Arda Uzunoğlu @aardauzunoglu]

Finally, we make data selection risk-controlled.

In practice, we often don’t know how noisy the target pool is. With a small calibration set, we can choose a trust threshold with statistical guidance, targeting a desired noise tolerance among selected weak labels.

[Arda Uzunoğlu @aardauzunoglu]

At least it was honest about it.

[Arda Uzunoğlu @aardauzunoglu]

Codex' goal feature is awesome. Today I asked it to do hyperparam search + dataset ablations for training a fasttext model until it has precision = 0.75 on the test set, and, after trying for an hour, it finally achieved it by training on the test set.

[Taiming Lu @TaimingLu]

Knowledge doesn't always flow downhill.

We find that in LLM pretraining, a weaker teacher can improve a stronger student, and pushing the teacher further can actually hurt.

New paper: Strong Teacher Not Needed? On Distillation in LLM Pretraining.

[Alex Dimakis @AlexGDimakis]

https://t.co/qqVULIZAWD