Mastodawn

Joanna Bryson, blathering May 27

First contributed talk I see @ #aamas q&a
q: re: teams & joint inteniton,does llm even understand those terms? A: our guidance does say shared awareness, informing others about abilities and limits.
I point out as next Q that the LLM will reply to that prompt by reporting the capabilities of the humans whose text makes up training data, not itself. Speaker: agents are only llm BASED
me: llm can't be the base, it's at best some kind of enhancement, the base has to have capabilities to report.

Joanna Bryson, blathering May 27

Robots are an important part of #aamas true from since back when I attended ATAL in the 1990s. Robot AI really has to work :-) The first keynote, Sabine Hauert
Designing Trustworthy Robot Swarms for Real-World Applications, was impressively focussed on making sure humans cound understand and even control emergent behaviour in real robot swarms e.g. for firefighting. Of course, if you can detect & bomb fire with water from the sky...so she needs to work on these things! #AAMAS

Joanna Bryson, blathering May 27

Similarly: Enhancing Goal Inference via Correction Timing
by Anjiabei Wang, Shuangge Wang, Tesca Fitzgerald

was a very nice robot paper about anticipating correction and seeing which aspects of it help with the behaviour, also relevant to human centring #aamas #AIEthics

Joanna Bryson, blathering May 27

From Actions to Words: Towards Abstractive-Textual Policy Summarization in RL
Sahar Admoni, Assaf Hallak, Yftah Ziser, Omer Ben-Porat, Ofra Amir

agent figures out which door the key opens.
environment observation creates agent actin impac ts policy. "typical RL" . What's new: use LLM to produce text describing these things / evnts, then use LLM to generate summaries, then evaluate to check precision (hallucinations) and recall

again, nice #humanCentring #aamas #aiethics

Joanna Bryson, blathering May 27

A bunch of people having trouble getting visas to come to theis EU but middle eastern country, – from the USA! Rice & Harvard! #aamas #trumpWars #WarsOfAggression

I got derailed checking I could give my keynote in Keynote, and then getting filmed as if I'd already given my talk (which I haven't actually written) so only caught half of

LEARN 7 – Social, Evolutionary and Swarm Learning

Surrogate-Augmented Deception in Reinforcement Learning (SAD-RL)
Joe Shymanski, Scott Nivison, Sandip Sen

Joanna Bryson, blathering May 27

looked like an excellent paper. itnereting question about when do you want to decieive? sad answer -- when you are in a conflict context. US SW authors weem very aware of that.

VGC-Bench: Towards Mastering Diverse Team Strategies in Competitive Pokémon
Cameron L. Angliss, Jiaxun Cui, Jiaheng Hu, Arrasy Rahman, and old friend Peter Stone!

Pokeman 10^136 they claim :-) violates every assumption chess etc use.

scraped 700K open team sheet battle logs #aamas #pokeman

Joanna Bryson, blathering May 27

ha ha looks like they are basically at intermediate human play which is great, but they won against an expert at the world level and had to respect their desire not to say how MUCH they won. Their method is not superhuman but they're still pleased given the challenging domain. More teams, per team performance falls but generalisation rises; all open source and online. #aamas Cameron Angliss presented online Jiaheng Hu is here doing q&a.

Joanna Bryson, blathering May 27

This is a massively parallel session event; when a speaker doesn't turn up without warning, then all the subsequent speakers suffer the loss of audience as well :-( #aamas #academicEthics

Joanna Bryson, blathering May 27

Can Vibe Coding Beat Graduate CS Students? An LLM vs. Human Coding Tournament on Market-driven Strategic Planning
Panayiotis Danassis, Naman Goel

No. People are fooled because benchmarks are binary. If you have agents play against each other, despite using most expensive LLM, lots of prompt strategies, LLM even degraded winning models when handed them as a part of the prompt! Runanable ≠ competitive. These were not great programmers, just msc students the AI can't beat. #aamas

Joanna Bryson, blathering May 27

We got a lot of stats in the opening (at 8:30am Cyprus, 7:30am Berlin!) 2/3 of the registrants are here, but it seems like more than half the talks are on line – seems like a lot of chinese & us researchers took advantage of this op and make up a lot of the papers. #aamas #hybrid

Joanna Bryson, blathering May 27

I've not attended mainstream AI conferences for some time, at least a decade, maybe ijcai2013 was the most recent. Observing some interesting and nice rituals deployed by many speakers:
1) "all of this code is online open source the links are in the paper and also here" often provided on two servers e.g. both github & huggingface
2) People are good & smooth about saying "that's a good question, that's a hard problem, we're open to suggestions" at the limits of the presented work. #aamas

Joanna Bryson, blathering May 27

OTOH some TERRIBLE practice wih people not turning up or just piping in badly with recorded talks. IMO there should be no recorded talks even if it is hybrid.

The final paper has FIVE authors present, and 1/4 the starting audience :-( #aamas

Joanna Bryson, blathering May 27

So the Victor Lesser award-winning dissertation couldn't get here due to visas. Congrats anyway Tonghan Wang, but it was too dense for me to live post. Now we are having the Silver Jubilee Panel, yay, they finally talk about my entry point: ATAL AAMAS was supposed to be about both single and multiagent levels of agent.

Joanna Bryson, blathering May 27

Pictures #AAMAS the names of the (oldest) people are in the next toot

Joanna Bryson, blathering May 27

Michael Wooldridge
Maria Gini and
Victor Lesser are the panel brought in to talk about this.
Oh, this is like the Distinguished Service Award at IJCAI; we're learning how great it is to be the ones who organise conferences. :-) #AAMAS

Joanna Bryson, blathering May 27

Ha, Lesser says ICMAS was named by Carl Hewitt because it was an AI Winter so that had doomed the Distributed AI meeting (1993! The year I started my PhD...) #aamas The FIRST one was 1978 Distributed Sensor Networks – then "agents" were called "noteds" but they still looked at interaction and cooperation. His own first paper was at Distributed Processing but that was very low level. Gini was the one that went from workshop to conference.

Joanna Bryson, blathering May 27

Wooldridge points out Lesser put the whole frst meeting on his visa card and was lucky a lot of peole came so he could pay his bill!

Wooldridge goes on to talk about #ATAL Everyone else was doing components, only hte robots were doing full systems, so that kicked off the whole arechitecture question, which he still thinks is the most relevant question in 2026. Oh, but now he's down the Bratman & Knowledge Representation (KR) rabbit hole (showing my biases :-) #scruffies #aamas

Joanna Bryson, blathering May 27

Wooldridge reminds us of how great #ATAL books were – we put in THREE versions of our papers, submission, revised after review draft, and third AFTER the event – you had to put in a corrected version that cited the other papers – became the biggest selling Springer LLNCS books as a consequence of his editorial work (I"m adding detail, Wooldridge is more modest.) #AAMAS The question was right. He was jealous of NIPS at one point, where's the Google cheque for us? :-D But now feels affirmed.

Joanna Bryson, blathering May 27

Gini – hey #AAMAS started in Bolognia, the first university of the world! That's why it went so well :-) History, food, we are humans, this has long-term impact.

She also makes an excellent point for the "future" question – #AAMAS needs to figure out how to reconcile what they are doing with what the #agentic systems are doing. I have tbh been disappointed by the lack of engagement to that here this year, but I expect Wooldridge's keynote will help tomorrow; we'll see.

Joanna Bryson, blathering May 27

Wooldridge says the first fully written by AI was genuinely MAS with agents playing different role, and it got in transparently to the scientific literature. He thinks that's astonishing, no one would have expected that in 2002. Nor that it could all be done in natural language NLP. #AAMAS he thinks there will be completely AI generated businesses in a couple of years (I hope not!!) I guess we'll have a human-centring #accountability fight on Friday after MY keynote.

Joanna Bryson, blathering May 27

#AAMAS medals for all the organisers from any year!!

Joanna Bryson, blathering May 27

Joanna Bryson, blathering May 28

Sven Koenig
ACM Autonomous Agents Award Talk

also going too fast for me to live toot sorry.

Joanna Bryson, blathering May 28

Sven Koenig
"I only work on cooperative robots because I build them. They have the team in mind. But sometimes we do decentralised"

genau :-) #humanCentring #aamas

Joanna Bryson, blathering May 28

Actually no demo but a great excuse to enjoy the weather. Oh, one‘s flying now.

#humanCentring #aamas

Joanna Bryson, blathering

TBH I went to an archealogical site during the coffee break after a quick look at posters, so missed the first half of the morning session. But enjoying an afternoon session now https://mastodon.social/@j2bryson.bsky[email protected]/116651971743987150n #AAMAS the session chairs are being more agile and fighting to keep audience, so you may not see talks in expected order but IMO building a community understanding is a higher proirity so I think this is good.

Joanna Bryson, blathering May 28

I talked about the hybrid problem a bit on my "main" account (IDK why more academic people are on bskky tthan here, but it seems to be true.)

https://mastodon.social/@j2bryson.bsky[email protected]/116646677254807157

Joanna Bryson, blathering May 28

Mike Wooldridge has been elected fellow of the Royal Society!

Joanna Bryson, blathering May 28

Wooldridge keynote #aamas