Justin Kang

bars

@HarryMayne5 @dswg97 Come check out our poster!

@dswg97 and I are presenting "A Positive Case for Faithfulness: LLM Self-Explanations Help Predict Model Behavior" (NSG): Today 2:00 – 3:45 PM, HALL A #3313

❓Our next BOLD paper at #ICML2026 is led by @HarryMayne5 and @Justinkangs. A Positive Case for Faithfulness: LLM Self-Explanations Help Predict Model Behavior When LLMs explain their decisions, can we trust those explanations? We present a new metric based on whether explanations help predict a model’s behaviour on similar inputs. We find self-explanations encode value information about model decision making! Give the paper a read: arxiv.org/pdf/2602.02639

A very strong team working on a great problem. Congrats @bneyshabur & Co!

@Suuraj Nice! Excited to read 😁

📢New ICML 2026 position paper: "Explainability research should prioritize foundations over ad-hoc methods." Despite countless explainability techniques (incl. feature attribution, concept-based methods, SAEs), explanations rarely influence real workflows. Why? We believe this is due to an over-emphasis on producing new methods instead of confronting core foundational questions, like definitions, properties, evaluations and utility. One example: despite 1000s of papers on SAEs, there is no rigorous, falsifiable definition of "concept", the very quantity it aims to discover. Without a definition, it is unclear what to evaluate, or how to quantify benefits. In the absence of gold-standard evals, papers are forced to use proxy evals (like sparsity-reconstruction tradeoffs for SAEs) that have weak links to the underlying task. How to break this cycle? The key idea is that clarity at any step: applications, evaluations, properties, or definition, helps define the other steps. They are all interdependent! Foundations doesn't necessarily mean epsilon-delta definitions, but greater clarity about goals. What we can do now: While writing the next XAI paper, we can ask ourselves whether we can partly address some foundational questions: better definitions, gold-standard evals, or well-motivated use-cases. We present a 5-point checklist in the paper as a useful starting point. Please read the paper for more detailed arguments and analysis! This was a great multi-institutional collaboration with an incredible set of colleagues: @ML_Theorist @lesiasemenova NaveFrost @CyrusRashtchian @valentynepii @ShichangZhang @hima_lakkaraju @CynthiaRudin @jennwvaughan Paper link: suraj-srinivas.github.io/pdfs/xai_posit… (Personal) Reflections: surajsrinivas.substack.com/p/explainabili…

@belindazli @PresItamar Cool to see NSG from our ICML 2026 Paper being used!

New paper, led by the amazing @PresItamar! Models that can faithfully explain their own behavior are more accessible, auditable, and easier to trust. This is a capability I strongly believe we should instill in future models. Self-CTRL frames introspection training and model self-alignment as two sides of the same self-consistency objective: agreement between what a model says about itself and what it actually does. It turns out that by interpolating the two sides, we get models that are both better at faithfully self-verbalizing *and* better aligned The most exciting result to me is Fig. 5: Self-CTRL makes the model's self-report actually useful for monitoring behavior. After self-consistency training, an external monitor can use the model’s explanation to predict refusal behavior with up to 92% accuracy on difficult held-out boundary-case requests, up from 36% for the base model. Check out Itamar's thread for more!

@DimitrisPapail Also having very little luck getting Fable to do basically any work remotely related to security or safety

Even if you tone it down it still routes you to Opus 4.8

As I suspected, ECHO will push the frontier of research towards RSI :)

@gdb Why are you telling us? Go make it better

Whenever I don’t use codex for a task, I ask myself why and usually realize that there’s some missing context, I needed to write a skill, or I just didn’t think to use it. Rarely is it because the task is outside of the capabilities of the model. Overhang right now feels large.

@jasondeanlee @roydanroy Been having tons of issues too

Whenever I try gemini deep think.

@melissapan @abeirami @AlexGDimakis Great discussion!

The fun continues 🔥 Now, we have our second panel with: @abeirami @AlexGDimakis and Mohammad Alizadeh Come by to hear more about hot takes from the field’s thought leaders!

🧵(1/8) An @OpenAI internal reasoning LLM achieved an AI Math milestone: solving an open problem central to its mathematical subfield— in this case, the unit distance problem of discrete geometry. We came across it in a side quest to truly push our model on the hardest problems.

The legend @PGelsinger has been vindicated!

I think the most beautiful application of information theory is "achievability" proofs. These prove that some engineering tasks is theoretically possible. Humans are an achievability proof for current AI. As we move beyond human capabilities this is no longer the case!

even if we do consider math as the pinnacle of human intelligence, it is not the kind of math that proves stuff, but the kinds of math that invents new concepts and then invents questions about them

@yoavgo Very good point!

@aminkarbasi This monograph is a great reference. I referred to it often when I was doing my work on formalizing some deep connections of group testing to data/feature attribution during my PhD.

Group testing is an elegant idea with deep connections to information theory, statistical physics, signal processing, etc.