Loading video...

Video Failed to Load

Go Home

A 7-million parameter model outperforming models a thousand times its size on tasks like ARC Prize. That's what recursive reasoning unlocks. In this episode of Decoded, YC's Ankit Gupta and Francois Chaubard break down two recent papers on recursive AI models, HRMs and TRMs, that are achieving state-of-the-art results...

128,225 views • 6 months ago •via X (Twitter)

29 Comments

Y Combinator's profile picture
Y Combinator6 months ago

Tune in:

dnu's profile picture
dnu6 months ago

@agupta @FrancoisChauba1 7M params is impressive, but recursive reasoning usually trades weights for more steps at inference. are you seeing a wall clock win too, or mostly benchmark gains once you let it think longer?

PARODY. The Dwark Side's profile picture
PARODY. The Dwark Side6 months ago

@agupta @FrancoisChauba1 Paper links: [2506.21734] Hierarchical Reasoning Model [2510.04871] Less is More: Recursive Reasoning with Tiny Networks

Dhruv Mishra's profile picture
Dhruv Mishra6 months ago

@agupta @FrancoisChauba1 Sounds like a fancy way of over fitting on a single thing, rather than anything novel.

SatyaNaaksh's profile picture
SatyaNaaksh6 months ago

@agupta @FrancoisChauba1 Does that mean that the model was better suited for this task because that is all it was trained for? Like, a specific key for a specific lock. Why would you go through the entire bundle of keys then?

ShadowFax's profile picture
ShadowFax6 months ago

@agupta @FrancoisChauba1 Does recursive reasoning at 7M params hold up on messy, ambiguous real-world inputs? ARC is impressive but structured. Enterprise data is chaos.

Freddy Snijder's profile picture
Freddy Snijder6 months ago

What is described as the HRM backprop method sounds very much like Truncated Backprop Through Time (BPTT), where backprop is continued, from a certain hidden state, representing an input sequence so far. This was used “back in the day” when we were training RNNs on long sequences.

AqibAi's profile picture
AqibAi6 months ago

This is a really important direction. If recursion at inference time can reliably replace raw parameter scaling for certain reasoning regimes, it reframes what “capability” actually means in LLMs. Still curious how well HRM/TRM generalize beyond curated benchmarks like ARC when you hit messy real-world distribution shifts—but the efficiency signal here is hard to ignore.

Teja's profile picture
Teja6 months ago

@agupta @FrancoisChauba1 recursive reasoning only helps

OpeningAi.com | For Sale's profile picture
OpeningAi.com | For Sale6 months ago

@agupta @FrancoisChauba1 This is the kind of efficiency breakthrough we actually need.

Adam Patarino's profile picture
Adam Patarino6 months ago

@paulg @agupta @FrancoisChauba1 Small, task focused models outperform large general purpose ones. The frontier of AI is in specialization.

Jacek (Jomsborg.eth)'s profile picture
Jacek (Jomsborg.eth)6 months ago

@agupta @FrancoisChauba1 I suppose its Looped Transformer

Crescitaly's profile picture
Crescitaly6 months ago

@agupta @FrancoisChauba1 A 7M parameter model outperforming ones 1000x its size on reasoning benchmarks is the clearest signal yet that scale isn’t the only path. Recursive computation lets small models think harder, not just be bigger. Architecture innovation is back.

Ghali Bennis's profile picture
Ghali Bennis6 months ago

the ARC numbers are wild. a 7M param TRM hitting ~45% on ARC-AGI-2 while frontier models with 1000x the weights sit in the single digits is the clearest signal yet that depth of compute at inference matters more than raw parameter count for actual reasoning. feels like the same lesson AlphaZero taught in 2017, just rediscovered for language.

valy's profile picture
valy6 months ago

@agupta @FrancoisChauba1 7M params doing ARC is the bigger story than scale laws this week

Launch Point Starbase's profile picture
Launch Point Starbase6 months ago

@agupta @FrancoisChauba1 7M params beating models a thousand times their size on ARC. The pure-scale thesis is aging poorly.

Adel Bucetta's profile picture
Adel Bucetta6 months ago

@agupta @FrancoisChauba1 the honest answer is that we're not even close to tapping into the real value of recursive reasoning yet. most papers are still just scratching the surface.

FlutterFlow's profile picture
FlutterFlow6 months ago

@agupta @FrancoisChauba1 woot woot!

ByteCrafter's profile picture
ByteCrafter6 months ago

@agupta @FrancoisChauba1 recursion replacing parameter count is the actual claim. the question worth watching is whether TRMs collapse on the same failure mode the giants did. ARC-AGI-3's biggest error bucket was wrong abstraction from training data and recursion alone doesn't obviously fix that one.

Hussain Hashim | Building SundayBack's profile picture
Hussain Hashim | Building SundayBack6 months ago

@agupta @FrancoisChauba1 @ycombinator that's wild. is it something to do with the way recursive models break down tasks? would love to know more about how they work.

nickster's profile picture
nickster6 months ago

@agupta @FrancoisChauba1 another cool interview by ycomb

Inflectiv AI ⧉'s profile picture
Inflectiv AI ⧉6 months ago

@agupta @FrancoisChauba1 This discussion on recursive reasoning highlights a promising path for efficient AI development. It shows how small models can outperform much larger ones on complex tasks.

GOLD TG's profile picture
GOLD TG6 months ago

@agupta @FrancoisChauba1 Hello Y combinator. I have an idea which I am implementing already. How can I share?

sabir hussain's profile picture
sabir hussain6 months ago

@agupta @FrancoisChauba1 Where do recursive models break down today?

Xzistor LAB's profile picture
Xzistor LAB6 months ago

@agupta @FrancoisChauba1

Atomreng ∞ Capital ⚛️'s profile picture
Atomreng ∞ Capital ⚛️6 months ago

@agupta @FrancoisChauba1

Today in AI's profile picture
Today in AI6 months ago

@agupta @FrancoisChauba1 Samsung TRM (arXiv:2510.04871), 7M params, hits 45% ARC-AGI-1 via 2-layer latent recursion , same score bracket as GPT-4o at 1000x the parameter count. Iterative z-loops beat transformer pass-throughs on fluid reasoning per FLOP.

Sandy.ETH's profile picture
Sandy.ETH6 months ago

@agupta @FrancoisChauba1 so answer wasn't bigger models just thinking longer

aitization 𝕏 's profile picture
aitization 𝕏 6 months ago

@agupta @FrancoisChauba1 YC is jumping on this nonsense train lol 😂

Related Videos

Small Language Models (SML) are the future of AI. "Small" (SML) instead of "Large" (LLM). These small models are highly specialized models with superhuman abilities on specific tasks. Here are two techniques to build these models: • Spectrum • Model Merging I give you a short introduction in the attached video, but here is a quick summary: Spectrum helps us identify the most relevant layers to solve one specific task. We can ignore everything else and focus on fine-tuning these layers. Using Spectrum, we can fine-tune models in a heartbeat. Model Merging combines multiple models into a unique, much better model than any of the individual input models. You can also combine models specialized in different tasks and get a model with multiple abilities. This is the state of the art of productizing models. It's what Arcee.ai's platform does behind the scenes. Arcee collaborated with me on this post and is sponsoring it. There are three main steps to produce a model for your particular use case: 1. You create a dataset by uploading your data. 2. You train a model. At this step, Arcee uses Spectrum and Model Merging to produce a highly specialized model for your task. 3. You can deploy that model to any environment you want. Three important notes: • Training process is 2x faster and 2x cheaper than regular fine-tuning. • Resultant models are smaller and have higher accuracy. • They create these specialized models from open-source models. Check this site so you can fully appreciate how this works: If you want to fine-tune an open-source model, consider Arcee's platform. This is the state of the art.

Santiago

164,162 views • 2 years ago

Thanksgiving-week treat: an epic conversation on Frontier AI with Lukasz Kaiser -co-author of “Attention Is All You Need” (Transformers) and leading research scientist at OpenAI working on GPT-5.1-era reasoning models. 00:00 – Cold open and intro 01:29 – “AI slowdown” vs a wild week of new frontier models 08:03 – Low-hanging fruit, infra, RL training and better data 11:39 – What is a reasoning model, in plain language 17:02 – Chain-of-thought and training the thinking process with RL 21:39 – Łukasz’s path: from logic and France to Google and Kurzweil 24:20 – Inside the Transformer story and what “attention” really means 28:42 – From Google Brain to OpenAI: culture, scale and GPUs 32:49 – What’s next for pre-training, GPUs and distillation 37:29 – Can we still understand these models? Circuits, sparsity and black boxes 39:42 – GPT-4 → GPT-5 → GPT-5.1: what actually changed 42:40 – Post-training, safety and teaching GPT-5.1 different tones 46:16 – How long should GPT-5.1 think? Reasoning tokens and jagged abilities 47:43 – The five-year-old’s dot puzzle that still breaks frontier models 52:22 – Generalization, child-like learning and whether reasoning is enough 53:48 – Beyond Transformers: ARC, LeCun’s ideas and multimodal bottlenecks 56:10 – GPT-5.1 Codex Max, long-running agents and compaction 1:00:06 – Will foundation models eat most apps? The translation analogy and trust 1:02:34 – What still needs to be solved, and where AI might go next

Matt Turck

168,243 views • 11 months ago