Lightnews — Scholar-powered news

spacecowboy @spacecowboy17.bsky.social · 2h

Bonus points for when the delta in technique is negative as with HRM -> TRM: arxiv.org/abs/2510.04871

Less is More: Recursive Reasoning with Tiny Networks

Hierarchical Reasoning Model (HRM) is a novel approach using two small neural networks recursing at different frequencies. This biologically inspired method beats Large Language models (LLMs) on hard ...

arxiv.org

spacecowboy @spacecowboy17.bsky.social · 2h

I should have googled it first: github.com/SamsungSAILM...

GitHub - SamsungSAILMontreal/TinyRecursiveModels

Contribute to SamsungSAILMontreal/TinyRecursiveModels development by creating an account on GitHub.

github.com

spacecowboy @spacecowboy17.bsky.social · 3h

@alexiajm.bsky.social any plans to release the source code?

When the HRM paper came out it was super helpful that they published the code for others to reproduce. Plus it helped the ARC folks to do the analysis.

1

spacecowboy @spacecowboy17.bsky.social · 3h

The author is on Bluesky!

Alexia Jolicoeur-Martineau @alexiajm.bsky.social · 25d

I'm working on super tiny neural networks that perform better than the big ones, it's kinda insane.

1 1

spacecowboy @spacecowboy17.bsky.social · 4h

Try For You - it shows content based on your likes. There is BG3, art content and a bunch more.

1

spacecowboy @spacecowboy17.bsky.social · 4h

Try For You - it has a muppets post at position #3. Lots of paintings.

This feed finds people who liked the same posts as you, and shows you what else they've liked recently.

1 1

spacecowboy @spacecowboy17.bsky.social · 6h

A good writeup! Very similar in the spirit to what I'm proposing. But instead of having the vetting of OK people as a separate process we get that for free from the same likes and flags.

spacecowboy @spacecowboy17.bsky.social · 15h

In For You I don't use the follow graph so I don't know if it performs worse.

We could use who you follow for the reply sorting: give a bigger (e.g., 10x) starting value of "like trust score" to your follows than to users that you don't follow.

spacecowboy @spacecowboy17.bsky.social · 16h

Yes, if we use only your dislikes and not other people's then it could be done privately on the client side.

It's a tradeoff between privacy and utility.

1 1

spacecowboy @spacecowboy17.bsky.social · 16h

Yes, and we can tune those parameters.

1

spacecowboy @spacecowboy17.bsky.social · 16h

Maybe we need a "choose your own algorithm" for ranking the reply section similar to custom feeds?

Then people can continue to use the standard client while testing different algorithms.

1 3

spacecowboy @spacecowboy17.bsky.social · 16h

We would need a third party (AppView) to have access to your and other people's dislikes.

1

spacecowboy @spacecowboy17.bsky.social · 16h

Regarding privacy, I think the dislikes should be private.

If we want to trust dislikes of other people then we wouldn't be able to do the reply sorting client-side - because your client would not have access to other people's dislikes.

2 1

spacecowboy @spacecowboy17.bsky.social · 17h

For You feed operates on the similar idea of giving more weight to users who liked the same posts as you.

It is not ideal - it may surface something completely random. For You is better though than a feed that sorts all Bluesky posts by the number of likes.

1 2

spacecowboy @spacecowboy17.bsky.social · 18h

Yes, the same user could be liking thoughtful replies on one topic (books) and trolling/inflammatory replies on another topic (sports).

When you like the former the algorithm would increase the visibility of the latter. But maybe it's ok because you may not read sports post replies.

1 1

spacecowboy @spacecowboy17.bsky.social · 1d

There are a few parameters to tweak - how much to increase/decrease the scores. This could be tested out.

4

spacecowboy @spacecowboy17.bsky.social · 1d

The starting conditions are: you have some nominal amount of "like trust" for every user and your "dislike trust" is 0 for every user. Which means that for a new user the reply section is sorted by like counts.

1 5

spacecowboy @spacecowboy17.bsky.social · 1d

Posted the idea: bsky.app/profile/did:...

spacecowboy @spacecowboy17.bsky.social · 1d

Here is an idea how to make the discourse healthier: change the incentive structure of the reply section.

Right now the replies are sorted by the number of likes. This is extremely gameable. 🧵

spacecowboy @spacecowboy17.bsky.social · 1d

3. as a secondary effect, posters/likers would know that their actions could have negative consequences for the amount of attention they will be getting in the future

1 6

spacecowboy @spacecowboy17.bsky.social · 1d

The result is that:
1. each user gets agency over how their future attention will be allocated in the reply section
2. we don’t have to resort to the blunt tool of blocking users

1 8

spacecowboy @spacecowboy17.bsky.social · 1d

When calculating the score of a reply we subtract from it how much you trust the dislikes of people who disliked this reply.

This may make this reply have a negative total score. In that case it can be hidden or grayed out.

1 5

spacecowboy @spacecowboy17.bsky.social · 1d

When you dislike a reply - you increase your "dislike trust score" for people who also disliked that post.

When you like a reply - you decrease how much you trust the dislikes of people who disliked this reply.

2 5

spacecowboy @spacecowboy17.bsky.social · 1d

Then when sorting the replies we sum the "like trust scores" of people who liked that reply.

There is another set of "trust scores" - how much you trust the dislikes of other people.

1 5

spacecowboy @spacecowboy17.bsky.social · 1d

Note: we treat the author of the reply as a liker.

We also need a dislike button (kind of like "show less like this" in feeds) - through which you can decrease your "trust score" for people who liked the content you didn’t like.

1 7

spacecowboy @spacecowboy17.bsky.social · 1d

We can fix that by creating a feedback loop that rewards good behavior and penalizes bad behavior. Who is the judge of deciding what is good vs bad, you ask? You are the judge.

When you like a reply - you increase your "trust score" toward people who liked that reply.

2 16