Alzbeta Manova
LightNews LightNews
banner
alzbetam.bsky.social
Alzbeta Manova
@alzbetam.bsky.social
160 followers 490 following 40 posts
🎓 PhD student @ University of Aberdeen | 📈 Math + 🧠 Psych + 🤖 ML.
🐕 Dog parent | 🧩 Puzzle solver
🏴󠁧󠁢󠁳󠁣󠁴󠁿🇨🇿🏳️‍🌈
Posts Replies Media Videos
alzbetam.bsky.social
Presented at the BMVA Symposium today!

Quick visit to London ✈️
Great feedback and ideas💡
And great food 🥐🍓
September 30, 2026 at 8:04 PM
Reposted by Alzbeta Manova
urielmenalled.bsky.social
Shout out to the #rstats {performance} package. Its check_model() function is super helpful for model evaluation.

#statistics #STEM
June 23, 2025 at 2:52 PM
Reposted by Alzbeta Manova
drsurekhadavies.bsky.social
Petition signable by UK citizens or residents, asking "the Government to introduce a statutory duty requiring AI developers to disclose the copyrighted works used to train generative AI in enough detail..."

💙📚 🗃 #academicsky
1/

petition.parliament.uk/petitions/76...
Petition: Protect creators' rights: require AI firms to disclose training data
We ask the Government to introduce a statutory duty requiring AI developers to disclose the copyrighted works used to train generative AI in enough detail for creators to identify use of their work, e...
petition.parliament.uk
August 6, 2026 at 6:10 PM
alzbetam.bsky.social
July in PhD:

🎤 Presented at a conference
💻 Collected data for a new experiment
🤖 Evaluated multiple VLMs
🎯 Got accepted to present at a symposium
🎂 Celebrated my birthday!
📊 Analyzed new data
🗓 planned the next 3 months of my PhD

#PhDSky #PhDLife
August 5, 2026 at 9:47 AM
Reposted by Alzbeta Manova
gracekind.net
This headline is extremely funny given what happened (OpenAI hacked HF by accident)

openai.com/index/huggin...
OpenAI and Hugging Face partner to address security incident during model evaluation
OpenAI and Hugging Face share early findings from a security incident during AI model evaluation, highlighting advanced cyber capabilities and lessons for defenders.
openai.com
July 21, 2026 at 8:16 PM
alzbetam.bsky.social
Nothing beats that feeling when your script and local server finally find each other and communicate correctly ✨️

Brings me back to my undergrad days full of debugging and codebase building 💻

#PhDLife #PhDSky #AIEvaluation #AI
Photo of 2 monitors. The left one shows Python code in PyCharm. The right one shows local server running in LM Studio.
July 29, 2026 at 3:25 PM
Reposted by Alzbeta Manova
gunnarblohm.bsky.social
New preprint from the lab asking what minimal parts of whole body posture/motion allow for rapid movement recognition

arxiv.org/abs/2607.13216
Classifying daily activities needs posture, reconstructing them needs motion
Humans recognize movements effortlessly, even from noisy and complex visual input. But what information in the stimulus allows humans to rapidly classify movements? No framework has systematically com...
arxiv.org
July 23, 2026 at 11:52 PM
Reposted by Alzbeta Manova
pardoguerra.bsky.social
This is quite the finding for a device thats just a plagiarism machine, or whatever we are calling it this week.

www.nature.com/articles/s41...
Large language models can predict the results of social science experiments - Nature
Large language models can be used to estimate the results of social science experiments about as accurately as a group of human forecasters—even for experiments published after the...
www.nature.com
July 9, 2026 at 1:25 AM
alzbetam.bsky.social
Day 2 of the Cyberpsychology Section Conference 🎓

Thank you to everyone who attended my talk on VLMs in action perception, loved all the feedback and discussions✨️

#cyber26 #conference #PhDSky
A photo of me standing behind the presenter's desk during my talk. In the background, the presentation title slide is visible, reading 'Obstacles to Intentionality, Divergent Interpretive Judgements of Hand Actions in Human and Artificial Observers'.
July 7, 2026 at 5:51 PM
alzbetam.bsky.social
Day 1 of the Cyberpsychology Section Annual Conference! 🧠🤖

Learned so much already and we are only half way through 🎓
#cyber26 #conference #PhDLife
A picture of my name badge for the conference. A poster (or a banner) of the conference name, links to the website and socials, and a sponsor list.
July 6, 2026 at 11:25 PM
Reposted by Alzbeta Manova
narijohnson.bsky.social
All AI evaluations embed assumptions about what "good" behavior looks like. Our #FAccT2026 paper explores how we can center the perspectives of impacted communities - specifically, subjects of AI-generated media - in designing LLM-as-a-judge evaluation rubrics.
A flyer teasing our research paper. The flyer contains a visualization of our approach: First, community members participate in creating an evaluation rubric, visualized as a list of criteria. The rubric is then given as input to an LLM judge.
June 24, 2026 at 6:31 PM
Reposted by Alzbeta Manova
ai-firehose.column.social
WorldOdysseyBench benchmarks interactive models, focusing on long-horizon stability in action following, visual quality, memory, and physics. This evaluation highlights gaps in existing models, improving the applicability of AI-generated environments. https://arxiv.org/abs/2606.31672
WorldOdysseyBench: An Open-World Benchmark for Long-Horizon Stability of Interactive World Models
ArXiv link for WorldOdysseyBench: An Open-World Benchmark for Long-Horizon Stability of Interactive World Models
arxiv.org
July 4, 2026 at 9:20 AM
alzbetam.bsky.social
4 years of growth, a 40-something cm chop, and lighter heads for both of us! ✂️

Officially a two-time hair donor 🙌 Highly recommend this haircut for the summer weather.

#HairDonation #DoGood
Side-by-side comparison of our hair before and after the donation.
June 30, 2026 at 4:27 PM
Reposted by Alzbeta Manova
earlkmiller.bsky.social
The brain continuously optimizing its predictions. Local "leaky" circuits correct errors, while hierarchies link these computations using nonlinearities shaped by prior expectations.

Predictive Coding with Bayesian Priors via Proximal Gradients
arxiv.org/abs/2606.08374
#neuroscience
Predictive Coding with Bayesian Priors via Proximal Gradients
We recast predictive coding as continuous-time proximal gradient descent applied to a regularized maximum-a-posteriori (MAP) objective. We study first a single-level problem and then a multi-level hie...
arxiv.org
June 23, 2026 at 7:00 PM
alzbetam.bsky.social
May in PhD:

✅Successful conference application
⚙️Organised school postgrad conference
📈Piloted 1 experiment
🎞️Processed 134 videos
📚Attended a teaching certificate workshop

Now off to a 3-week road trip! 🚗
#PhDSky #PhDLife
May 31, 2026 at 6:04 PM
Reposted by Alzbeta Manova
craignewmark.bsky.social
AI can't handle the truth:

"AI is still getting things wrong, more confidently than ever"

www.axios.com/2026/05/30/a...
AI can't handle the truth
Confident chatbots could encourage users to stop fact-checking.
www.axios.com
May 30, 2026 at 1:11 PM
alzbetam.bsky.social
Helped organise & presented at our annual school postgrad conference last week!
Sitting at the tech desk on stage definitely helped with the stage fright when sharing my research in action perception!

Huge thanks to organisers, attendees & presenters ✨
#Conference #Psychology #PhDLife #PhDSky
Me on stage presenting my project "Drivers behind Perceptual Bias in Observed Hand Actions". The presentation shows the title, authors, organisaitons and is in green, white and yellow colours. Photo of my behind the scenes set-up as the tech admin at the conference. It shows a monitor with a presentation loaded, and the lecture theatre is visible in the background. Photo of the organisation committee on the stage at the end of conference. The committee consists of 6 PhD reserachers.
May 29, 2026 at 3:56 PM
Reposted by Alzbeta Manova
ai-firehose.column.social
ProcBench shifts focus from final outcomes to execution-process quality in LLM coding agents, revealing defects and enhancing reliability and governability. This method uncovers flaws often missed by traditional evaluations. https://arxiv.org/abs/2605.20251
ProcBench: Evaluating Process-Level Defects and Control Preservation in LLM Coding Agents
ArXiv link for ProcBench: Evaluating Process-Level Defects and Control Preservation in LLM Coding Agents
arxiv.org
May 23, 2026 at 8:40 AM
alzbetam.bsky.social
Always love a good MLLM evaluation paper. Over half (51%) of "correct" model deductions about video subjects are ungrounded shortcuts rather than real cue processing!
#AI #MLLM
Perception or Prejudice: Can MLLMs Go Beyond First Impressions of Personality?
Multimodal Large Language Models (MLLMs) are increasingly deployed in human-facing roles where personality perception is critical, yet existing benchmarks evaluate this capability solely on numerical ...
arxiv.org
May 22, 2026 at 12:07 PM
Reposted by Alzbeta Manova
privacyinternational.org
AI can reveal your location from a single photo. Vision‑Language Models can determine where any photo is taken, without any GPS data

What does that means for your privacy?
Nowhere to Hide? Privacy Risks and Policy Implications of AI Geolocation
One of the most surprising — and concerning — capabilities of the newest Artificial Intelligence (AI) systems is their ability to infer geographic location from images. Vision‑Language Models (VLMs)...
privacyinternational.org
May 16, 2026 at 5:15 PM
Reposted by Alzbeta Manova
markriedl.bsky.social
ArXiV has a new LLM policy

(Screenshots with alt text so you don’t have to click through to the other place and see all the stupid responses)
Attention @arxiv authors: Our Code of Conduct states that by signing your name as an author of a paper, each author takes full responsibility for all its contents, irrespective of how the contents were generated. 1/ If generative AI tools generate inappropriate language, plagiarized content, biased content, errors, mistakes, incorrect references, or misleading content, and that output is included in scientific works, it is the responsibility of the author(s). 2/ We have recently clarified our penalties for this. If a submission contains incontrovertible evidence that the authors did not check the results of LLM generation, this means we can't trust anything in the paper. 3/ The penalty is a 1-year ban from arXiv followed by the requirement that subsequent arXiv submissions must first be accepted at a reputable peer-reviewed venue. 4/ Examples of incontrovertible evidence: hallucinated references, meta-comments from the LLM ("here is a 200 word summary; would you like me to make any changes?"; "the data in this table is illustrative, fill it in with the real numbers from your experiments") end/
May 14, 2026 at 9:50 PM
alzbetam.bsky.social
5 days until our school postgrad conference! 🎓

Throwback to last year when I was just a guest and didn't realize that organizing the thing is slightly more stressful than just presenting 😅

#ConferencePrep #PhDLife #PhDSky
A photo of me presenting from a podium at last year's conference. A slide is projected behind me showing the start of my presentation title: "Action Perception and AI Generated Content: Perceived Intentionality of AI Agents and Drivers behind Perceptual Bias".
May 13, 2026 at 9:09 AM
Reposted by Alzbeta Manova
visionaryvistas.bsky.social
A 1 a.m. thought for the CogSci undergrads:
Everyone talks about the "cool" theories, but no one tells you that Mathematics is what actually builds your future. Go for the math, even if it’s not required. Your future self will thank you for the analytical edge. 🧠📉 #ScienceSky#AcademicSky
May 8, 2026 at 8:02 PM
alzbetam.bsky.social
Nothing cures research procrastination like a conference acceptance notification 😅

I suddenly have all the motivation to complete my study 📊
#PhDLife #PhDSky #Cyber26
May 9, 2026 at 7:53 PM
Reposted by Alzbeta Manova
mensmachina.com
Truly excellent work from Anthropic. I think this is a promising interpretability technique, but what I really like is that they’ve made code available for open weight models.
Natural Language Autoencoders
Turning Claude's thoughts into text
www.anthropic.com
May 7, 2026 at 10:03 PM