#GPT5 is STILL having a severe confirmation bias like prev SOTA models! 😜. Try yourselves (images, prompts avail in 1 click):. It's fast to test for such biases in images. Similar biases should still exist in non-image domains as well.

RT @mrnuu: @anh_ng8 @taesiri @an_vo12 GPT5 Pro, the most advanced version, reasoned for 1 minute 19 seconds and came up with the wrong answ….

RT @PetarV_93: ok! part 2 of my early-stage ai research lore - ask and you shall receive!. so now i found myself in a research group as a f….

RT @2prime_PKU: Anyone knows adam?

@grok . How many chess pieces are there on this board?

@grok How many points are there on the star in the logo of this car?

RT @gdb: We've published a position paper, with many across the industry, calling for work on chain-of-thought faithfulness. This is an opp….

RT @Cohere_Labs: Supported by one of our grants, @an_vo12, Mohammad Reza Taesiri, and @anh_ng8 from @kaist_ai, tackled bias in LLMs. Their….

RT @an_vo12: 🚨 Our latest work shows that SOTA VLMs (o3, o4-mini, Sonnet, Gemini Pro) fail at counting legs due to bias⁉️. See simple cases….

RT @MichLieben: This isn't hallucination in the traditional sense. Grok's math was nearly correct. But it confidently applied PhD-level t….

RT @savvyRL: The opportunity gap in AI is more striking than ever. We talk way too much about those receiving $100M or whatever for their j….

RT @giangnguyen2412: Today I finished my PhD at @AuburnEngineers with @anh_ng8 . What’s next? Off to @guidelabsai to build foundation AI….

RT @samim: That @cursor_ai silently downgrades the working model from Claude4 to Claude3.5 during an active coding session, is borderline c….

Asking GPT-4o to generate images in the style of @DiegoCusano_ .shows an existing gap between the real Cusano vs. GPT-4o in creativity / wittiness. (random samples) .Sometimes it signs "Cusano" at the bottom.

RT @yacineMTB: I got fired today. I'm not sure why, I personally don't think there is a reason, or that it's important. When I joined twi….

2015 in Boston: I attended my first #CVPR (1 paper in the main conference). 2025 in Nashville: my 3rd grader has his first #CVPR (no workshop or conference papers though 😜). #CVPR2025

TAB is currently applied to 2-image image difference captioning. But could one apply it to general VLMs? 🙂. Work led by the amazing ☘ @Pooyanrg (catch him at #CVPR2025 ).w/ Hung Nguyen, @savvyRL , Long Mai. Paper: Code:

🌟 Bonus: Training VLMs with TAB on image-difference captioning over 3 datasets (MS COCO, CLEVR-Change, Spot-the-Diff) also, consistently improve captioning performance over the baseline (MHSA). Less -> More!

Similarly, not letting VLMs to see anything by zero-ing out all patch attention (i.e. all attention mass now is on the CLS token) causes VLMs to be *see no changes* and therefore outputs "there is no change" ✅.

Editing the bottleneck attention causes the attention to change, which flips the incorrect caption from window [❌] to wheel [✅], which is correct. → empirical causal relationships between VLM attention and text outputs.