Influential Papers

The title is all you need! Super cool!

Makes a really interesting case for disambiguating prompt 'engineering' from a possible 'prompt science'. The writing is very compelling. A helpful analogy presented in the paper is the idea that plant breeders were able to infer a lot of the internal structure of plants before genetic theory explained it. Not sure where I stand on this still but it's given me a lot to think about, especially w.r.t. 'aesthetic' concerns in ML research that bar 'prompting' from being seen as a legitimate research method. I do agree with the paper's claim that many important works in NLP are basically interfaces/structures upon prompting, and we shouldn't be afraid to more closely associate them with a 'prompt science'.

Really useful description of AI as a governance system, akin to markets, democracy, and bureaucracy -- a lens that lets political scientists to usefully contribute towards thinking on AI.

Advocates for rhetoric over composition as the methodology for writing pedagogy in the AI era.

Really nice argumentative piece on why we can build much better AI interfaces than chat interfaces.

An ambitious yet informed vision for what it would mean and cost for us to have *convivial* tools -- tools that we can make and shape our own lives for joy and purpose.

Interesting thoughts on what kind of methodology suits interpretability as a growing, immature field.

Swanton reads Nietzsche's 'immoralism' and 'egoism' as articulating virtues of the 'mature egoist' -- someone who has overcome immature egoism (ressentiment, self-deception, reactive self-assertion) in favor of genuine self-affirmation and creative engagement with the world. Setting aside questions of fidelity to Nietzsche's texts, I find this a compelling moral outlook: a rejection of both slavish self-denial and petty self-aggrandizement in favor of something like joyful, active, honest self-cultivation.

Super interesting and illuminating perspective explaining why supposedly deep-learning-unique phenomena like deep double descent, overparametrization, etc. can be explained using soft inductive biases and existing generalization frameworks. The references are a treasure trove!

Advocates for understanding LLM behavior as indicative of nuances in human social behavior.

Kontorovich reflects on how AI and formal verification systems like Lean are reshaping mathematical practice, intended for ICM 2026. What I find compelling is that the examples come from someone deeply embedded in both traditional research mathematics and these new technologies. The discussion of the 'Bitter Lesson' applied to mathematics is sobering, and the concrete examples of what formal verification looks like in practice are clarifying. A useful document for thinking about what 'doing math' might mean going forward.

Identifies four cognitive behaviors (verification, backtracking, subgoal setting, backward chaining) that predict whether a model can self-improve via RL. The key finding is striking: it's the presence of reasoning behaviors, not answer correctness, that matters. Models exposed to training data with proper reasoning patterns -- even incorrect answers -- matched the improvement of models that had these behaviors naturally. A useful framing for thinking about what 'reasoning' actually is in these systems.

A method to probe structure in language models without any notion of ground truth, relying instead on the consistency property of tru statements.

Useful outline of what HCI researchers can contribute to 'AGI'. It's not obvious (and people may fear that) interaction problems will be solved by AGI. Perhaps not?

A recognition that not only must AI 'align to' human (values, behavior, knowledge, etc.) (-- whatever this means), but we also need to think about how humans might 'align' to AI by working with AI-structured systems. This paper recognizes social 'looping effects' brought about by AI and its behavior.

Interesting way to automatically label datasets using mutual predictability and logical consistency.

A compelling articulation of what human-AI communication could look like. Proposes neologism learning.

By modeling individual views rather than an aggregated 'view', we can explicitly define the voices 'heard' in making a decision and consider counterfactuals.