Lightnews — Scholar-powered news

Andrea Pugnana @andrepugni.bsky.social · 21d

For other believers, link to the old arxiv, arxiv.org/abs/2503.161.... updated version to be out soon.

Deferring Concept Bottleneck Models: Learning to Defer Interventions to Inaccurate Experts

Concept Bottleneck Models (CBMs) are machine learning models that improve interpretability by grounding their predictions on human-understandable concepts, allowing for targeted interventions in their...

arxiv.org

Andrea Pugnana @andrepugni.bsky.social · 21d

Deferring Concept Bottleneck Models go to
@neuripsconf.bsky.social
2025! Thanks to all my amazing coauthors (especially @rmassidda.it who is the only one on 🦋 as far as I know) who believed in the marriage between CBMs and Learning to Defer! See u in San Diego! 😄🌊🏖️

1 1 3

Reposted by Andrea Pugnana

Emanuele Marconato @ema-ridopoco.bsky.social · May 3

Now in Thailand to present our paper at #AISTATS2025 🇹🇭🌴

📍Today at 3:00-6:00 pm, poster number 118!

More details here:

openreview.net/forum?id=XCm...

All or None: Identifiable Linear Properties of Next-Token...

We analyze identifiability as a possible explanation for the ubiquity of linear properties across language models, such as the vector difference between the representations of “easy” and “easiest”...

openreview.net

1 2

Andrea Pugnana @andrepugni.bsky.social · Apr 29

👀 Want to dive deeper? Curious how ML + humans really perform together?
📍Come find us at our poster: Hall A-E 160, May 3rd.

5/5

Andrea Pugnana @andrepugni.bsky.social · Apr 29

🔍 We define wo scenarios:
1️⃣ Both human and ML predictions for deferred cases? We can estimate individual & aggregate causal effects.
2️⃣ Only human predictions for deferred cases? We can use Regression Discontinuity Design for local causal effects.

4/5

1

Andrea Pugnana @andrepugni.bsky.social · Apr 29

In our work, we bring causal inference to the arena. Using the potential outcomes framework, we ask: What’s the actual impact of deferring to humans on system performance? 🎯

3/5

1

Andrea Pugnana @andrepugni.bsky.social · Apr 29

🧠🤖 Deferring systems let ML models pass the mic to humans when they are better at predicting. But what is the actual impact on predictive performance of such systems? Surprisingly, such a question is still overlooked.

2/5

1

Andrea Pugnana @andrepugni.bsky.social · Apr 29

On my way to Mai Khao🏝️ for #AISTATS2025, where I will present A Causal Framework for Evaluating Deferring Systems, a joint work with Filippo Palomba, Jose Alvarez and Salvatore Ruggieri.

1/5

1 1 2

Andrea Pugnana @andrepugni.bsky.social · Apr 29

👀 Want to dive deeper? Curious how ML + humans really perform together?
📍Come find us at our poster: Hall A-E 160, May 3rd.

5/5

Andrea Pugnana @andrepugni.bsky.social · Apr 29

🔍 We define wo scenarios:
1️⃣ Both human and ML predictions for deferred cases? We can estimate individual & aggregate causal effects.
2️⃣ Only human predictions for deferred cases? We can use Regression Discontinuity Design for local causal effects.

4/5

1

Andrea Pugnana @andrepugni.bsky.social · Apr 29

In our work, we bring causal inference to the arena. Using the potential outcomes framework, we ask: What’s the actual impact of deferring to humans on system performance? 🎯

3/5

1

Andrea Pugnana @andrepugni.bsky.social · Apr 29

🧠🤖 Deferring systems let ML models pass the mic to humans when they are better at predicting. But what is the actual impact on predictive performance of such systems? Surprisingly, such a question is still overlooked.

2/5

1

Reposted by Andrea Pugnana

Steve Azzolin @steveazzolin.bsky.social · Apr 25

Happening tomorrow!

Poster number 508
Saturday's session 10-12:30

Steve Azzolin @steveazzolin.bsky.social · Apr 17

Faithfulness of GNN explanations isn’t one-size-fits-all🧢
Our last @iclr-conf.bsky.social paper breaks it down across:

1. Evaluation metrics
2. Model implementations
3. OOD generalisation

w: Antonio L. @looselycorrect.bsky.social @andreapasserini.bsky.social

1/5

1

Reposted by Andrea Pugnana

Steve Azzolin @steveazzolin.bsky.social · Apr 17

Faithfulness of GNN explanations isn’t one-size-fits-all🧢
Our last @iclr-conf.bsky.social paper breaks it down across:

1. Evaluation metrics
2. Model implementations
3. OOD generalisation

w: Antonio L. @looselycorrect.bsky.social @andreapasserini.bsky.social

1/5

1 3 5

Reposted by Andrea Pugnana

Causal Club @handle.invalid · Apr 10

If you are into Structural Causal Models, you're going to love Deep SCMs! ❤️

On the next (and last for now!) talk, Audrey Poinsot (INRIA) will review DSCMs, their applications and open challenges.

🤌 See you tomorrow April 11th at 16:00 (CET) at meet.google.com/tng-aibh-yuz

🗓️ tinyurl.com/deepscm

2 3

Andrea Pugnana @andrepugni.bsky.social · Apr 5

❗️Deadline extended 18 April❗️
Working on Human-AI interaction and trustworthy AI? Consider submitting your ongoing work to our workshop tcai2025.isti.cnr.it @hhaiconference.bsky.social
Co-hosted by Apoorva Singh, Bruno Lepri, @burcusayin.bsky.social and Roberto Pellungrini in Pisa.
#AI #HHAI

TCAI2025

tcai2025.isti.cnr.it

2 2

Andrea Pugnana @andrepugni.bsky.social · Mar 27

Great talk tomorrow at the @causalclub.di.unipi.it by Nina Corvelo Benz on counterfactual token generation.

Causal Club @handle.invalid · Mar 27

Sure! Here's your post to advertise the next Causal Club talk:

🤖 Tomorrow, Nina Corvelo Benz from MPI-SWS will present her work on counterfactual token generation in LLMs!

🤌 See you in Pisa on March 28 at 16:00 (CET) or live at meet.google.com/tng-aibh-yuz

🗓️ Add it here: tinyurl.com/ctgllms

Reposted by Andrea Pugnana

Causal Club @handle.invalid · Mar 20

Time for the next Causal Club speaker!

⏳ Mouad El Bouchattaoui (Paris-Saclay University) will present a novel method for conterfactual regression over time (openreview.net/pdf?id=bKOZY...).

🤌 See you in Pisa on March 21 at 16:00 (CET) or live at meet.google.com/tng-aibh-yuz

2 4

Andrea Pugnana @andrepugni.bsky.social · Mar 12

Also quite curious to see what new technical issue will appear on Friday.

1

Andrea Pugnana @andrepugni.bsky.social · Mar 12

Happy to share our second episode with @giovannidetoni.bsky.social ready to rock at @causalclub.di.unipi.it on Friday at 16:00.

Riccardo @rmassidda.it · Mar 12

The second appointment of the year for the Causal Club will feature algorithmic recourse, time, causality, and pizza*! 🍕

See you on Friday! 🎉

*slices for in-presence participants only, as we are still waiting for food teleportation through Google Meet

Causal Club @handle.invalid · Mar 12

Still thinking about that denied loan?

🏦 Then you can't miss @giovannidetoni.bsky.social talk on the causal role of time in algorithmic recourse

🤌 See you in Pisa on March 14 at 16:00 (CET) or live at meet.google.com/tng-aibh-yuz

🗓️ Click here to add it to your calendar: tinyurl.com/timecausalclub

1

Reposted by Andrea Pugnana

Causal Club @handle.invalid · Mar 7

⏰ Don't miss Mátyás talk today at 3PM!

🎥 See you online meet.google.com/cqt-ufji-xfz

🤌 ...or live in the CS Department of the University of Pisa!

Mátyás Schubert @matyasch.bsky.social · Mar 4

A few weeks ago, I presented SNAP at the wonderful #Bellairs Workshop on Causality in Barbados🐢

This Friday 🫰meets 🤌 as I will get to present SNAP again at the kick-off of the newest season of @causalclub.di.unipi.it! Check out this, and their other amazing upcoming talks at causalclub.di.unipi.it

3 7

Andrea Pugnana @andrepugni.bsky.social · Feb 28

The Causal Club is back for Season 3! Do not miss it!

Causal Club @handle.invalid · Feb 28

Here's a talk you can't miss!

🔨 Next Friday, @matyasch.bsky.social will discuss how to efficiently estimate causal effects on large unknown graphs.

🤌 See you in Pisa on March 7 at 15:00 (CET) or live at meet.google.com/cqt-ufji-xfz

🗓️ Click here to add it to your calendar: tinyurl.com/snapcausclub

Mátyás Schubert @matyasch.bsky.social · Feb 13

Do you want to estimate causal effects for a small set of target variables without knowing the causal graph, but discovering it takes too long? Now you can get adjustment sets in a SNAP🫰accepted at #aistats2025!

📜 arxiv.org/abs/2502.07857
🧩 matyasch.github.io/snap/
🧵 1/10

Andrea Pugnana @andrepugni.bsky.social · Feb 25

Are you in Philly for @realaaai.bsky.social 2025, and still do not know what oral to attend? Do not worry. Also, ML models do not know many things. Check what they know they don't know with Salvatore Ruggieri, who will present our joint work in the senior track morning session on Friday!

2 4

Reposted by Andrea Pugnana

Luca Sofri @sofri.it · Feb 19

DOGE Claimed It Saved $8 Billion in One Contract. It Was Actually $8 Million.
www.nytimes.com/2025/02/18/u...

DOGE Claimed It Saved $8 Billion in One Contract. It Was Actually $8 Million. (Gift Article)

The biggest single line item on the website of Elon Musk’s cost-cutting team appears to include an error.

www.nytimes.com

2 3 41

Andrea Pugnana @andrepugni.bsky.social · Jan 22

Finally, we thank the reviewers for their valuable suggestions and will add the link to the complete version as soon as possible. See you in Mai Khao! 4/4