Ayush Mittal

Goodhart Resistant!
Co-founder & CEO at what actually moves the needle.

[Ayush Mittal @ayushmi5]

What world models actually need for agentic systems?

Before calling something a world model for agents, ask these five questions:

1. State - Is the latent a sufficient statistic of history, or is loop closure faked with caches?
2. Data - Were the training actions exogenous, or chosen by a policy that saw the state?
3. Interventions - Does it answer do(a), or only “what followed a in the log”?
4. Counterfactuals - Can it hold the world fixed and replay a different action?
5. Cost - Does the plan survive when experiments aren���t free and there is no retry?

Video prediction, scaled forever, answers none of these by itself.

Prediction is rung one. Agents live on rung three.

That gap is what we are solving for.

[Ayush Mittal @ayushmi5]

We believe recursive self-improvement is being framed too narrowly.

Self-modification is a capability.

RSI begins when a system can keep improving the process that improves it, while the meaning of "better" remains well-defined under that self-reference.

Building towards Constructive Gödel Machine:

The learner should improve itself.
The improver should rewrite its own rule.
The improver rewriting the improver.

Not an LLM, no agent, we believe exact arithmetic, priced experiments, every proof through Lean's kernel, nothing tuned after the draw will be the true proof of RSI.

[Noumenal AI @noumenal_world]

Accepted across #COLM and multiple #NeurIPS workshops

Our research program: infrastructure for recursive self-improvement.

Measurement that survives optimization.
Generalization that survives self-rewrite. Counterfactuals under intervention.
Formal learning theory that actually compiles.

@ayushmi5 @Zetetic_Dhruv

[Ayush Mittal @ayushmi5]

Two identical lines: same equipment, same feed, same spec. They don't produce at the same cost, and the delta persists month after month. What differs is the thousand small decisions per shift. Cost per tonne is set by the variance of decisions, not the equipment.

[Tanishq Mathew Abraham, Ph.D. @iScienceLuvr]

Music-JEPA: Learning a World Model of Sound from Action

"we propose to learn a world model of piano sound using JEPA by framing music as an action-conditioned system: the audio is treated as the state, and the pianoroll as the instrument action."

"Experiments show that the learned model captures the relationships between musical actions and their resulting sound. The resulting representations support downstream tasks, including beat tracking, composer identification, and key estimation, and enable piano transcription via planning, by searching for actions that best explain a target sound."

[Daniel Jeffries @Dan_Jeffries1]

Catalog of incorrect AI predictions.

Doomers from the rationalist cosplay community are left out, presumably because their predictions have a near 100% failure rate.

When people proudly proclaim that we need to 'act now' or make laws because something is inevitable, point them at this list and tap it loudly because it ain't.

https://t.co/VDVJeECxDr

[Ayush Mittal @ayushmi5]

A plant historian is deep in time and narrow in action: twenty years of data, mostly one policy doing what works. A game log is the opposite — mistakes are free, so players explore everything. Every game log is a natural experiment. Every plant log is a habit.

[Ayush Mittal @ayushmi5]

Milton Friedman's thermostat, industrial edition: a good operator compensates every disturbance, so outcomes barely move and regression concludes the inputs don't matter. The better the control, the flatter the data. Good control destroys the evidence of its own necessity.

[Ayush Mittal @ayushmi5]

Action-conditioning feels causal. But when the policy chose the action from the state, the action carries information about unobserved state - so the model can use it as a sensor instead of a cause. do(a) needs a to vary independently of what produced it. Logs rarely oblige.

[Ayush Mittal @ayushmi5]

Two different worlds run by two different policies can write identical logs. That's the problem inside next-state prediction: the data is a product of world x policy, and nothing in the loss factorizes the product. Low error certifies the composite, not the world.

[Ayush Mittal @ayushmi5]

You can't A/B test a blast furnace. Nearly everything special about physical-world AI follows from that sentence: no resets, no cheap exploration, bounded trial, irreversible error. Retry is a property of the environment, not the model - and most of ML grew up where retry was free.

[Ayush Mittal @ayushmi5]

The first section hides a complete epistemology: a new ontology is new physics only if it changes an observable. ML inverts this. Two models can match every logged observable and still be different theories. One learned the world, one learned the policy that generated the log. Physics breaks ties by asking nature new questions. A frozen dataset can't ask anything. So "empirically equivalent" ends the argument in physics and begins the problem in machine learning.

[Ayush Mittal @ayushmi5]

A world model can be photoreal and causally empty. Fidelity measures how well you reproduce what happened. Identifiability is whether the model's answer to "what if I act" can be certified at all. We benchmark the first. Acting needs the second. Fidelity is not identifiability.

[Ayush Mittal @ayushmi5]

This is the most beautiful result on ARC-AGI-3 yet, the detail that matters more than the number for me is that: [schema] didn't predict its way to saturation, it experimented its way there. No rules, no reward means mechanism must be discovered, and discovery runs through intervention: hypothesise, act, backtest, edit the program. Objects first, then what your actions do to them. Nouns, then verbs. The one thing ARC-AGI-3 still gives away free is the interventions.

[Haven Feng ✈️ CoRL @HavenFeng]

Today, we’re introducing [schema]: a harness reaching 99% RHAE with Opus 4.8 + Fable 5 and 95.35% with GPT-5.6 Sol on ARC-AGI-3 Public set.

[schema] makes an LLM think like a physicist. 🧵

[Ayush Mittal @ayushmi5]

I think the split reduces to two numbers. Sufficiency gap I(future, history | latent): here nonzero means the encoder isn't an information state, and loop closure has to be faked with caches. Lipschitz constant of the closed-loop latent dynamics: above 1, rollout error compounds geometrically, and you build error banks. Consistency and stability are measurements of these two, runtime is amortisation. Prosthetics appear exactly where a theorem was skipped.

[Ayush Mittal @ayushmi5]

L_curv is a soft Koopman constraint: you're asking for coordinates where the dynamics run straight exactly the regime where MPC turns benign. We have done latent curvature spikes which in our case coincide with contact events on Push-T. It's cheap though align per-step latent curvature against contact onset over the rollouts, and if the spikes land on contacts, your residual 64% is physics, not optimisation.

[Haiyu Wu @HaiyuWu1]

Today's leading SSL methods rely heavily on training heuristics—EMA, teacher-student training, layer freezing, and more—to remain stable.

Regularization-based methods are much simpler, but have long failed to match the generalization performance of leading SSL approaches.

VISReg breaks this trade-off.

We introduce VISReg, the first regularization-based SSL method to achieve state-of-the-art out-of-domain generalization, outperforming MoCov3, DINO, data2vec, iBOT, I-JEPA, MAE, and DINOv2.

arXiv: https://t.co/rq687ag2XV
Github: https://t.co/XeD8TZroFg

[clem 🤗 @ClementDelangue]

or maybe let's make sure the most important technology in the history of humanity is not controled by just 4 men? Aka let's push for open science & open-source AI to distribute capabilities, power and wealth!

[Sriram Krishnan @sriramk]

It is clear open source models and harnesses are having a moment. There's a few factors at work

1/ It is now obvious that you can catch up to near-SOTA performance and do so with a clear training lineage. See:@thinkymachines Inkling launch today.

2/ There are several well-funded, talented teams building open weight models now in the US and abroad. Along with the explosing of other near SOTA models (Grok/Cursor, Muse Spark), it is clear we are going to have a diverse ecosystem of models atleast on coding and agentic use.

3/ Organizations are increasingly looking for control over how their data is used and are willing to trade off some access to frontier level tokens for this control. Organizations and countries are increasingly nervous about the frontier labs potentially competing with them down the road and don't want their data to enable a future competitor.

4/ Open source is a slider: you could bring your own open harness, your evals, your business context and are free to pick and choose your model of choice.

5/ Companies have now actively shifted from "how do we get our people to use tokens" to being uncomfortable with their token cost ballooning without a clear line to revenue.

6/ Geo-politically, countries will be weighing open weight models as a way to get frontier-level tokens inside controlled environments that may not be otherwise possible.

All of this leads to more choice for all of us !

[Ayush Mittal @ayushmi5]

@ramez The mandatory gate fails for a deeper reason than innovation cost: passing a chosen battery certifies behavior on that battery, not in deployment. For systems that act, risk lives on the distribution the agent picks once out. Tests can falsify safety, they can’t verify it.

[Ayush Mittal @ayushmi5]

https://t.co/GLyFSZGfLA