Short notes after COLM 2026

Agents, Everywhere

  • The level of abstraction keeps getting higher every year: individual capabilities (e.g., classification) → autonomous agents → collective intelligence.

  • Multi-agent orchestration is already a thing. Now people are pushing toward simulating individuals and even entire societies.

  • I’m not sure how people assess progress in this space beyond benchmark performance.

User Modeling and Social Simulation

  • I’ve become more interested in cognitive science-adjacent research, so I spent some time going through posters and attended the user modeling and simulation workshop.

  • Lots of heuristics. We don’t really have a good theory of how human societies emerged and evolved—they just happened naturally. Perhaps that’s why we don’t have great intuitions for simulating them either. At least people are building synthetic datasets based on social situations and personas, then trying to match human behavior. Maybe that’s a reasonable way to make progress in this space.

  • I think the goals and expectations in this space should be stated more explicitly. Group behavior depends heavily on who’s in the group, and matching the behavior of one group doesn’t necessarily mean a synthetic data pipeline generalizes well.

    • (A bit of self-reflection here, since I’ve also used similarity to a particular human group as an evaluation metric in a recent project.)

  • We need more discussion about what the desiderata should be, but I’m not sure the current conference format encourages that kind of discussion.

    • I like Trends in Cognitive Sciences in that sense.

The Diversity Problem

  • A prevalent view seems to be that RL kills diversity.

  • But diversity is exactly what we need for things like social simulation, automated scientists, and LLM creativity. How do we preserve it while optimizing for performance? I guess we’ll see a bunch of papers on this in the next conference cycle…

Measuring Superintelligence

  • François Chollet’s panel discussion at the Frontier Data Summit was pretty interesting. One point that stood out to me was how to benchmark superintelligence. As many existing benchmarks are saturating, perhaps we should start thinking about how to evaluate capabilities beyond human levels.

  • I guess this will become an active research area, since no one really knows what the “super” in superintelligence should mean.

Benchmarking everywhere

  • The ML community seems to be converging on a familiar recipe: build harder benchmarks, max them out, and call it progress.

  • I guess we should be skeptical of benchmarks, but the community seems to be addressing this skepticism by scaling up benchmarking efforts. No benchmark is perfect, but perhaps the idea is that having more benchmarks helps compensate for their individual limitations.

  • People are building benchmarks for almost every aspect of everyday life. Why don’t we just record everything and turn it into benchmarks lol.

  • Benchmarking seems to be becoming a more accessible way to make research contributions in academia, especially compared to training models.

    • Another avenue is interpretability…

On the model/architecture side

  • Not my area of expertise, but diffusion language models seem to be gaining traction.

  • I didn’t see many papers on alternative architectures like Mamba or hybrid models (similar to last year, I guess?). Perhaps that’s because developing and training new architectures is much more resource-intensive than other types of research.

Research in the Age of AI Anxiety

  • The fear: OMG, AI is pushing humans out of the research loop.

  • It feels like everyone is churning out papers faster than ever.

  • My coping mechanism: if researchers are doomed, society is probably collectively doomed anyway. At least we won’t be alone on our way to doom, lol.

  • Some personal (and potentially controversial) thoughts:

    • I don’t see anything inherently wrong with publishing AI-assisted research, as long as it produces genuine insights. Most problematic cases seem to come down to sloppy research rather than AI itself. I like John Locke’s quote: “…the gold is nevertheless gold, and will enrich the man that employs his pains to seek and separate it.” — Of the Conduct of the Understanding (1706), §3.

    • One major problem is that the current conference review system isn’t designed for AI-assisted knowledge production at scale. The number of reviewers is limited, while AI can generate research outputs at an ever-increasing rate. Naturally, there’s growing pressure to scale up reviewing with AI as well, sometimes without much transparency.

    • But why does this matter? I think the deeper problem is that our evaluation of research has always been noisy. Humans make imperfect value judgments, and models trained on human judgments inherit those imperfections. Researchers have always had incentives to game evaluation systems, but AI might enable reward hacking at an unprecedented scale.

    • I guess market forces will address some of these problems on the industry side. Research that makes people’s lives easier or more entertaining will probably find its way into products anyway.

    • Academia is trickier. We value research even when it doesn’t directly generate revenue, but how do we determine what’s worth pursuing? One possible answer is its potential implications for industry—not necessarily immediate commercial value. I don’t have a clear answer here, though.

    • Values are certainly shifting. For example, being a mathematical genius might not be regarded with the same reverence as before. I think much of that prestige came from the rarity of such capabilities, and the increasing availability of intelligence could significantly reshape what society values.

    • Conferences are becoming weird places. As the research landscape gets noisier, I wonder if we’ll see a return to authority-based knowledge: people trusting work from major institutions or companies rather than reading individual papers and judging their value themselves—or simply relying on what AI considers valuable (I guess we’ll see papers analyzing AI models’ value systems, lol). You know, too many papers and not enough time to read them properly.

  • Anyway, what a chaotic time to be a researcher. Hope we all survive.