#ARCPrize
LightNews LightNews
xchatter.techmeme.com
This tweet appeared under this Techmeme headline:

@arcprize:

Gemini 3 Flash Preview (High) on ARC-AGI Semi-Private Eval - ARC-AGI-1: 84.7%, $0.17/task - ARC-AGI-2: 33.6%, $0.23/task Competitive performance at a substantially lower cost than other frontier models [image]
December 17, 2025 at 5:52 PM
nathanlabenz.bsky.social
@TheZvi for Live Player analysis, a p(doom) update, and what's virtuous to do now

@GregKamradt of @arcprize on LLMs' incredible efficiency gains, his favorite creative approaches of 2025, and what 2026 winners might look like
December 15, 2025 at 8:51 PM
sama.zpravobot.news.ap.brid.gy
Sam Altman 𝕏📝💬 https://xcancel.com/arcprize:
390x cost reduction in a year!
https://xcancel.com/sama/status/1999191411313508704
December 11, 2025 at 7:02 PM
stenichele.bsky.social
Our paper "ARC-NCA: Towards Developmental Solutions to the Abstraction & Reasoning Corpus" has been awarded a Runner-Up #ARCprize

Congrats team: E. Guichard, F. Reimers, M. Kvalsund, M. Lepperød, & me

Thanks F. Chollet, G. Kamradt, M. Knoop!

Paper: etimush.github.io/ARC_NCA/
December 5, 2025 at 9:12 PM
pekka.bsky.social
ARC-AGI is probably the most overrated and misleadingly marketed benchmark and the ARC Prize Foundation must be in denial of all its issues if they don't understand why their apples to oranges comparisons do not align with their expectations based on very misleadingly reported human baselines.
ARC Prize @arcprize Nov 18 Frontier AI reasoning systems are now closing the complexity scaling gap between ARC-AGI-1 and ARC-AGI-2 This is surprising, as these same systems also make obvious mistakes on easy tasks (for humans) from ARC-AGI-1. We're not sure why and invite help from the community to study this phenomenon Full solution logs are linked in last tweet ARC Prize @arcprize For example, ARC-AGI-1 Public Eval task http://arcprize.org/play?task=14754a24 This task involves completing cross shapes and is very intuitive for humans, while Gemini 3 Deep Think misses the nature of the task on both attempts
November 22, 2025 at 9:54 PM
poetiq-ai.bsky.social
Is more intelligence always more expensive? Not necessarily.

Introducing Poetiq. We’ve established a new SOTA and Pareto frontier on @arcprize using Gemini 3 and GPT-5.1.
November 20, 2025 at 6:21 PM
handle.invalid
@rohanpaul_ai https://x.com/rohanpaul_ai/status/1990925140604170428 #x-rohanpaul_ai

Quote: https://x.com/arcprize/status/1990820655411909018

💸 Gemini 3.0 Pro and especially Gemini 3 Deep Think just jumped to the top of the ARC-AGI reasoning leaderboard with much higher scores than ...
November 18, 2025 at 11:45 PM
xchatter.techmeme.com
This tweet appeared under this Techmeme headline:

@arcprize:

Gemini 3 models from @Google @GoogleDeepMind have made a significant 2X SOTA jump on ARC-AGI-2 (Semi-Private Eval) Gemini 3 Pro: 31.11%, $0.81/task Gemini 3 Deep Think (Preview): 45.14%, $77.16/task [image]
November 18, 2025 at 5:14 PM
altryne.bsky.social
Evals are nuts 📈

With a massive 1501 @arena (+50 on 2.5 pro), 91.9% on GPQA diamond and 37.5% on HLE, this is the most powerful LLM intelligence we've had access to yet.

And w/ DeepThink + tools an even more impressive 41.1% on HLE and an unprecedented 45.1% on @arcprize
November 18, 2025 at 4:00 PM
timkellogg.me
notes from ARC

lots of open questions. i personally am holding my excitement until they get answered
ARC Prize & @arcprize • 19h Our notes: - TRM has a higher runtime than HRM even though it is smaller. Our hypothesis is this is due back propagation happening across all steps, whereas HRM only did partial steps Open question: Is TRM better because it is smarter? or because it trains for longer? If you used fixed-compute for both, would performance be the same? ••• ARC Prize » @arcprize • 19h - Is TRM similarly robust to number of augmentations as HRM was? - Switching from linear layers to attention is interesting, attention performed worse on a smaller task. Why? It might be computationally less efficient, but why is it that much worse on Maze? ARC Prize & @arcprize • 19h Our call for the community: Split Pre-training and inference in TRM Currently pre-training and inference are coupled in TRM. Additional batches of tasks need to be pre-trained again. This augmented TRM would likely be able to run on Kaggle for ARC Prize 2025
October 17, 2025 at 12:26 PM
timkellogg.me
ARC reproduced TRM

i’ve been skeptical of this one, a 7M that beats o3-pro. bold claims!

ARC-AGI repro isn’t everything. Even HRM was reproduced, but HRM was also found to not be interesting

huggingface.co/arcprize/trm...
arcprize/trm_arc_prize_verification · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
huggingface.co
October 17, 2025 at 12:26 PM
handle.invalid
@rohanpaul_ai https://x.com/rohanpaul_ai/status/1976351262183653776 #x-rohanpaul_ai

Quote: https://x.com/arcprize/status/1976329182893441209

GPT-5 Pro now holds the highest verified frontier LLM score on ARC-AGI’s Semi-Private benchmark 👏

It still lags the OG o3-preview model that...
October 9, 2025 at 6:45 PM
xchatter.techmeme.com
This tweet appeared under this Techmeme headline:

Mark Kretschmann / @mark_k:

Potentially huge AI breakthrough: "Less is More: Recursive Reasoning with Tiny Networks" A 7B model that scored very highly on the @arcprize benchmark. Francois Chollet called it "impressive work". [image]
October 9, 2025 at 10:53 AM
bytetrending.bsky.social
Arcprize: Win Prizes & Boost Your Business

Unlock exclusive opportunities! Discover how arcprize is reshaping digital rewards and offering innovative ways to engage audiences. Learn about its unique approach & explore what makes it a rising star in the prize-linked savings space.
Arcprize: Win Prizes & Boost Your Business
Unlock exclusive opportunities! Discover how arcprize is reshaping digital rewards and offering innovative ways to engage audiences. Learn about its unique approach & explore what makes it a rising star in the prize-linked savings space.
bytetrending.com
September 23, 2025 at 7:28 PM
altryne.bsky.social
These weeks are getting denser and denser!

We covered @openai Codex, ICPC updates and Usage paper, @Meta display glasses, @reve incredible UI and Model, @LumaLabsAI HDR Ray3, chatted with @jerber888 about @arcprize SOTA with Grok, @theworldlabs Marble and SO MUCH MORE 👇
September 19, 2025 at 12:45 AM
sciencecqfd.bsky.social
ARC-AGI The General Intelligence Benchmark tinyurl.com/23cxfqc3 via #ARCPrize #ScienceCQFD
September 11, 2025 at 2:28 PM
pekka.bsky.social
Summary by Lauren Wagner, who is "Building trustworthy AI @arcprize".
Lauren Wagner @typewriters The team @arcprize analyzed @makingAGI's paper, claiming a new kind of AI model (HRM) - inspired by neuroscience - delivers unprecedented reasoning power on complex tasks, without pretraining or CoT. By scoring on hidden tasks and running ablations, @arcprize found that performance *didn't come from* the new hierarchical architecture - modeled on how the brain processes information - but from how it re-checks and improves its answers several times before finalizing them. This is a great example of how testing on ARC can decode AI progress and cut through hype. I'm excited about neurotech, but this isn't really an example of neuroscience accelerating AI progress. The model learns each evaluation task really well, rather than generalizing to new ones. Gains mostly come from iterative refinement and tailored training, rather than a new reasoning architecture. A regular transformer, trained in the same way, would come close to its performance.
August 16, 2025 at 9:57 AM
handle.invalid
@rohanpaul_ai https://x.com/rohanpaul_ai/status/1948556806147428689 #x-rohanpaul_ai

Quote: https://x.com/arcprize/status/1948453132184494471

So Qwen3-235b Instruct officially gets 11% on ARC-AGI-1 and 1.3% on ARC-AGI-2 (semi-private sets).

The key thing is at about $0.003–$0.004 pe...
July 25, 2025 at 2:01 AM
justinhjohnson.com
What do you think – how soon will AI close the gap? Share your takes, follow for more AI benchmark insights, and join the discussion. #ARCAGI3 #AGI @arcprize @fchollet
🧵 4/4
July 20, 2025 at 4:01 PM
mm-hacker-news.bsky.social
Grok 4 (Thinking) achieves new SOTA on ARC-AGI-2 (X)
https://twitter.com/arcprize/status/1943168950763950555
July 10, 2025 at 10:35 AM
xchatter.techmeme.com
This tweet appeared under this Techmeme headline:

@arcprize:

Grok 4 (Thinking) achieves new SOTA on ARC-AGI-2 with 15.9% This nearly doubles the previous commercial SOTA and tops the current Kaggle competition SOTA [image]
July 10, 2025 at 5:43 AM
handle.invalid
@swyx https://x.com/swyx/status/1941634409934790788 #x-swyx

Quote: https://x.com/aiDotEngineer/status/1941632491292590235

Connections is a fun word game that is still unsolvable for top reasoning models. It's like the @arcprize AGI benchmark, except that it a...
July 5, 2025 at 11:30 PM
fry69.altq.net
And -> xcancel.com/scaling01/st...
Lisan al Gaib @scaling01 16h my minions returned and they observed no performance difference between o3 versions ARC Prize @arcprize 16h After the o3 price reduction, we retested the o3-2025-04-16 model on ARC-AGI to determine whether its performance had changed. We compared the retest results with the original results and observed no difference in performance. Jun 11, 2025 · 4:35 PM UTC
June 12, 2025 at 9:19 AM
fry69.altq.net
See also -> xcancel.com/jeremyphowar...
Jeremy Howard retweeted Greg Kamradt @GregKamradt 16h After all the comments around OpenAI possibly serving a new model or distilled model we retested the data to check tldr: no performance change was noticed We reran o3 against our original results in April and saw the same perf Price reduction is real! ARC Prize @arcprize 16h After the o3 price reduction, we retested the o3-2025-04-16 model on ARC-AGI to determine whether its performance had changed. We compared the retest results with the original results and observed no difference in performance. Jun 11, 2025 · 4:27 PM UTC
June 12, 2025 at 9:19 AM
timkellogg.me
despite accusations that OpenAI did something fishy to get the 80% price drop, it appears they’re serious when they said nothing changed
ARC Prize @arcprize Follow After the o3 price reduction, we retested the 03-2025-04-16 model on ARC-AGI to determine whether its performance had changed. We compared the retest results with the original results and observed no difference in performance. ARC Prize @arcprize • 7h This indicates that there is no evidence of a distilled or altered model being served after the
June 12, 2025 at 12:08 AM