Influential Papers

Really fascinating to explore all of these features. I am personally curious about 'nontrivial' features which we don't quite have great words to describe succintly / aren't just simple (possibly fuzzy) substring searches.

Language modeling in terms of natural language predicates

Challenges the assumption that implicit measures are superior to self-reports in psychological research. Argues that self-reports demonstrate greater reliability, stronger predictive validity for both deliberate and spontaneous behaviors, and unmatched flexibility in exploring complex psychological constructs.

Cool classic HCI-style exploration into a new interaction for code editing using visual sketching, interpreted by VLMs.

Creates a DSL and program optimizer to apply LLMs to large/complex document processing. Provides a more structured and visible way to interact with LLMs on large text corpuses.

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.

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

This paper uses sparse autoencoder features to identify possible hypotheses to explain relationships between text and a dependent variable.

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

Advocates and proposes directions for systematic AI evalutions involving real human interaction.

Argues that human-AI collaborations represent a continuation of our basic nature to build hybrid thinking systems that fluidly incorporate non-biological resources. Discusses 'extended cognitive hygiene' as essential for critically evaluating what we incorporate into our digitally extended minds.

By learning explicitly defined concepts to bridge independent and dependent variables, we can better interpret model decisions and intervene on mistakes.

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.

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?

Really nice technical details and knowledge on training and understanding large sparse autoencoders,

Makes what I think is an important point that there is no such thing as 'eliminating bias' in AI, and gives reasonable measures for how to proceed nevertheless in not building partisan hack LLMs. I personally find the 'Output Transparency', 'System Transaprency', and 'Neutrality Through Diversity' approximation techniques the most compelling. I appreciate what this paper does for broadening a notion of 'neutrality' in AI, drawing on the large existing body of political and philosophical work making the point.

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

A comprehensive historical survey of the development of abstract algebra from its origins to modern times.