International Conference on Learning Representations

So one obvious direction is, is scaling up.

Just across the three main conferences, uh, last year, so this is ICML, ICLR, and NeurIPS, there was something like 12,000 papers accepted.

So even if we just say, "Okay, we want to, uh, just take a couple of years' worth of papers," we could, uh, increase the size of our task set by, uh, two orders of magnitude.

Uh, and that of course, would enable us to train a larger model as, as a coding agent and hopefully get an even stronger improvement.

They are multi-embodiment, basically.

So on the research background of scaling law and robotics, there is like an ICLR paper from 2025 called Data Scaling Laws in Imitation Learning for Robotic Manipulation.

that fitted power roles over environments, objects, and demonstrations.

Also, Xiaomi Robotics, in July, they published 100,000 hours of embodiment-free UMI trajectories.

For concreteness, this might look like.

This would be a funny world to live in, wouldn't it? Imagine the ICLR paper titles.

Subheading.

For dot B.

But how do the authors investigate this potential reward hacking?

They construct a controlled corpus from 120 anonymized ICLR 2026 submissions, resulting in 4,200 full-paper manuscripts that are rewritten with changes in six rhetorical dimensions.

These dimensions are claim and novelty stance, scope and generalization, quantitative evidence framing, contribution structure, technical register and formalism, and lexical and syntactic complexity.

And how do they measure the impact of these different rhetorical choices on AI reviewers?

How did the authors actually investigate this potential reward hacking in AI reviewers?

Evan, they employed a very structured approach for this investigation.

Firstly, they constructed a validated baseline for comparison by using 120 anonymized full-paper manuscripts from actual ICLR 2026 submissions.

These were selected to cover a broad quality range, ensuring a diverse dataset for analysis.