Fazl Barez

Let's build AI's we can trust!

[Fazl Barez @FazlBarez]

Really grateful to have 7 papers accepted at @icmlconf 2026, including 2 spotlights!

Massive thanks to all my collaborators—I’ve been lucky to work with such brilliant people

See you all in Seoul?…it feels surreal saying that while still in Rio for ICLR 😄 #ICML2026

[Fazl Barez @FazlBarez]

Wonderful. Now it’s officially blessed by the second-highest authority. The highest signed off in 1883!

[Christiane Amanpour @amanpour]

Barcelona’s wondrous church La Sagrada Familia came one step closer to a long-sought completion this week! Gaudi’s architectural marvel got a blessing from Pope Leo XIV to inaugurate its latest, and tallest, tower.

[Fazl Barez @FazlBarez]

@BlancheMinerva Im still surprised, I expected approx 30% online with previous years afair?

[Fazl Barez @FazlBarez]

@TheAndiPenguin Why only MTS?

[Fazl Barez @FazlBarez]

@sarahwiegreffe @icmlconf Ohhh no I’m tempted to go! What papers you thinking of?

[Stella Biderman @BlancheMinerva]

In film, "we'll fix it in post" is what you say when something went wrong on set and you don't want to redo it. AI research has made it our entire methodology: train the model, then patch whatever comes out. Our new ICML oral argues this can't be the basis of a science of AI. 🧵

[Stella Biderman @BlancheMinerva]

Read the full paper: https://t.co/gdqpVQ08Ur or come listen to our oral @icmlconf!

Huge thanks to my co-authors @aflah02101 @niloofar_mire @linguist_cat @FazlBarez @nsaphra

Stay tuned for a related workshop (hopefully) at NeurIPS too!

[Fazl Barez @FazlBarez]

@bneyshabur Congrats and great goal!

[Dongrui Liu @dong_rui39501]

[1/8]
🐶 AgentDoG 1.5: a diagnostic guardrail for AI agents.
Most safety models say "unsafe." AgentDoG names which step, what risk, what harm.
4B hits 55.2% on fine-grained diagnosis. GPT-5.4: 25.8%.
📄 https://t.co/oOGDYfEHjV
💻 https://t.co/wmi1DUtyV7
#AISafety #Agent

[Fazl Barez @FazlBarez]

Really glad to see @_clementneo doing great work! Proud moment

[Neo Research @NeoResearchAI]

We're Neo Research (新衡). Asia’s first independent frontier AI safety evaluation & research lab.

Today we're publishing our first report: an independent safety evaluation of DeepSeek v4 Pro. (1/5)

[Fazl Barez @FazlBarez]

@brianchristian @CHAI_Berkeley Congrats! 🙌🏼

[Fazl Barez @FazlBarez]

@mattbotvinick @AnthropicAI Congrats!

[Lucas Irwin @lucasjamesirwin]

Really exciting to see politicians picking up on our recent work on token taxes! (full paper to be released soon) @FazlBarez @Tung-Yu Wu @icmlconf

[Fazl Barez @FazlBarez]

The ultimate question is whether these boundaries are permanent and human nature only, or if they are just skills AI hasn't mastered ...yet?

[Pope Leo XIV @Pontifex]

Artificial intelligences do not undergo experiences, do not possess a body, do not feel joy or pain, do not mature through relationships, and do not know from within what love, work, friendship or responsibility mean. Nor do they have a moral conscience, since they do not judge good and evil, grasp the ultimate meaning of situations, or bear responsibility for consequences. They may imitate or even simulate, but they do not understand what they produce, for they lack the affective, relational, and spiritual perspective through which human beings grow in wisdom. #MagnificaHumanitas

[Fazl Barez @FazlBarez]

@ShayneRedford @alex_pentland @sarahookr @PeterHndrsn Congrats

[Fazl Barez @FazlBarez]

Suppose evaluations were sufficient for deployment: What’d happens once models become continuous learners?

Simple: every update would require new evaluation and your pre deployment evals would expire at step 1!

Checkout our new paper lead by @LPacchiardi

[Lorenzo Pacchiardi @LPacchiardi]

🚨 New paper: AI evaluation is structurally unsuitable for continual learning (CL). To address this, evaluation should be centred on the "behavioural trajectories" that CL systems develop, with the goals of characterising possible behaviours and forecasting their evolution. 🧵

[Fazl Barez @FazlBarez]

@DakingRai @ZiyuYao @megamor2 Our works been public since Sep'25 and accepted at ICML, it's great to see this used to refine tasks. raw single-instance gradients are noisy, using an un-smoothed s exposes your clusters to noise. BoN would directly stabilise DCD. Looking forward to the update and discussion! 4

[Fazl Barez @FazlBarez]

@DakingRai @ZiyuYao @megamor2 Also, evaluating specialised subsets hits the exact instance-level evaluation bottlenecks we solved which is why DCD’s CMD′ metric (clipping over-recovery to zero) exactly mirrors the bounding principles of our introduced NDF metric.
(3/)