Basis @ ICML 2026 — outreach field sheet

Your thesis framing -- world models, user models, and self models as three facets of the same problem -- is basically the theory-of-mind side of what Basis is trying to build for MARA's embodied agents. I'd love to hear how you'd think about 'self models' for a system that has to

Between your robust-Bayesian-optimization-under-corruption work and now co-founding Recursive to build AI that experiments on itself to accumulate knowledge, you're running a strikingly similar bet to MARA's — active experimentation as the core engine of discovery. What convinced

Diffusion Policy kicked off generative visuomotor control, and your Structured 4D Latent Predictive Model (ICML 2026 poster) pushes that into planning over learned 4D world state -- close to the loop MARA needs for a robot building and revising its own abstractions through intera

Your ICML 2026 paper with Fan Feng on task-sufficient world models — closing the loop between agentic exploration and structured model learning — is close to word-for-word how we frame the active-experimentation problem in MARA; curious how identifiability guarantees carry over o

Your AutoToM/MindZero line treats theory-of-mind as automated model-based inference over agent programs rather than a fixed cognitive module -- curious how you handle search-over-model-space when the agent population is open-ended, which is exactly MARA's problem.

RAP uses the LM itself as both world model and planner doing MCTS over it -- we want MARA's world models to have their abstractions actively discovered through experimentation rather than fixed at pretraining; what breaks first when you push RAP-style planning into an open-ended

CausalGame is testing whether LLM agents can actually do causal reasoning under intervention rather than just pattern-match correlational text -- that's exactly the gap Basis is trying to close with MARA's active-experimentation loop. Curious whether your PC/FCI-style discovery m

Your trajectory-balance/continuous-GFlowNet line and the new reinforced-SMC amortized sampler (your ICML 2026 poster) is exactly the learned-proposal-for-intractable-posteriors work ChiRho's inference stack runs into at scale; curious how you think about amortizing inference acro

Navigation World Models and the parallel stochastic-gradient planning work (your ICML 2026 poster) both push on using a learned world model directly for control -- the same problem MARA attacks from the program-synthesis side with explicit compositional abstractions; curious wher

CauSciBench and CLadder probe the same causal-reasoning gap in LLMs that motivates AutumnBench at Basis — curious whether the benchmark design generalizes to embodied/interactive settings the way MARA needs.

PySR treats scientific law discovery as program synthesis over symbolic expression trees -- MARA wants the analogous thing for compositional world models discovered via robot experimentation rather than curve-fitting; where does your tractability analysis for symbolic regression

pykan hit 16k stars in weeks because people want interpretable, symbolic-flavored networks -- and 'From Kepler to Newton' asks exactly what MARA cares about: what has to be built into a world model so it discovers the actual force law instead of a trajectory-fitting shortcut. Did

Your position paper frames world models as the intermediary layer between agents and reality -- close to how Basis thinks about MARA's embodied model discovery; curious how WorldGym or World-Gymnast's RL-in-a-world-model loop would generalize to an agent that's also choosing whic

Your Calibrated Test-Time Guidance work gets honest posteriors out of guidance-based samplers at inference time -- similar to what Basis needs from its inference stack so a MARA agent knows how uncertain it is about a hypothesis before choosing its next experiment.

Temporal Straightening finds the straight-line path in latent space instead of unrolling a generative rollout step by step -- does that hold up once the environment has genuinely discontinuous causal structure, or does it need an explicit abstraction layer, which is the bet MARA

Your amortized-inference/BayesFlow line and Basis's simulation-based approach to world models are solving the same problem from different sides -- I'd love to compare notes on where invertible-network SBI breaks down versus where an active-experimentation agent like MARA could ju

You were a scikit-learn core dev before you built SAIL -- Basis is trying to do something similar with MARA: build a robot that runs its own simulation-based-inference loop against the physical world instead of a synthetic likelihood. Appa's approach to global-scale probabilistic

Asta's push toward full-cycle agentic scientific discovery is exactly the terrain MARA is staking out for embodied, world-model-driven science - where do you see the ceiling for language-only hypothesis generation without an actively interactable world model?

Probabilistic numerics treats computation itself as inference - close to how we think about ChiRho's effect handlers as a substrate for treating simulators as probabilistic objects. Does the ProbNum toolkit have a story yet for solvers over discrete or branching structure, rather

AutumnBench's world models ultimately need posterior inference over combinatorial, non-continuous program structure - a much harder summary-statistic problem than the continuous population-genetics settings ABC was built for. Does your minimum-distance-summaries approach to robus

Your Test-of-Time adaptive-submodularity and safe-BO work is the theoretical backbone of active experimentation MARA needs; curious how your ICML 2026 BO-for-policy-search poster connects to closed-loop world-model discovery rather than static black-box optimization.

Your optimization-centric generalized-Bayes framework looks like exactly the robustification Basis's PPL stack needs once the 'model' is a program synthesized on the fly by an LLM agent rather than a fixed parametric family — have you thought about generalized Bayes over posterio

Your Bayesian Learning Rule paper is basically the unifying algorithmic lens we want for probabilistic-programming-native optimizers - curious how you'd map ChiRho's causal effect-handler abstraction onto a BLR instance.

Your Modern Bayesian Experimental Design synthesis and the BED-LLM line (using LLMs for intelligent information-gathering under a Bayesian design objective) map closely onto how Basis's MARA agents need to actively choose experiments/queries to reduce uncertainty about a world mo

Your flow-matching calibration-under-misspecification work is directly relevant to how we validate causal/world models in ChiRho - when the simulator itself might be wrong, how do you disentangle inference error from model misspecification in practice, and would that change how y

Path-dependent Discrete Amortized Inference and your torchgfn library are squarely in the amortized-inference space Basis leans on for MARA's world-model posterior estimation — curious how you think GFlowNets' path-based credit assignment could plug into program-synthesis-style d

Zergham's bilevel-planning world-model synthesis work with you and Josh feels like a close cousin of what Basis's MARA project is trying to do — programs that discover and compose abstract world models through active experimentation rather than being handed a fixed representation

Your microservices root-cause-discovery paper is a concrete instance of the causal-discovery-under-intervention problem MARA hits constantly -- a robot figuring out which of its own actions caused a failure; curious how the soft-intervention identifiability results generalize to

Aether AI's whole premise — causality-empowered world models for VLA — is basically the bet MARA is making too, just from the founder side instead of the lab side; curious what made you go the startup route instead of staying purely inside HDSI.

Kernel two-sample test and causal representation learning connect directly to Basis's ChiRho causal world-model discovery work — curious what inductive bias for compositional causal structure looks like at MARA's scale.

Your Hilbert-coresets and 'automated Bayesian inference' agenda is basically the theoretical backbone underneath what we're trying to make practical and turnkey at Basis in our own PPL work.

Your constrained Bayesian experimental design work is close to the active-experimentation loop MARA needs for robots choosing which physical experiment to run next — curious how BED assumptions break down once the environment pushes back on the action space.

TheoryCoder synthesizes causal theories as forward simulators for bilevel planning -- MARA is trying the analogous thing for robots discovering compositional world models through interaction; where does program synthesis diverge from RL trial-and-error once the simulator has to b

Centaur's attempt at a single foundation model of human cognition, plus your curiosity/exploration work in children, both get at what MARA is chasing on the robotics side -- models that structure their own exploration the way people do; curious how curiosity-driven exploration in

Your invariant-prediction and anchor-regression line of work is basically the theoretical backbone we lean on when we ask which parts of a learned world model should stay fixed under an intervention — curious how you'd apply invariance tests inside an active-experimentation loop

BayesAgent's idea of verbalizing probabilistic graphical models inside an LLM's reasoning loop is close to how we think about giving MARA agents an explicit, inspectable world model rather than a black-box chain of thought — curious how you see that scaling from static PGM verbal

MLGym and the Meta AI Scientist work both frame research itself as a scored task — MARA wants agents that design the experiments that generate new benchmarks, not just search existing ones. Where does open-endedness break down when the environment is a physical robot, not a proce

Your SMC-based approach to stochastic optimisation fits how we think about scalable inference for probabilistic programs at Basis -- does the particle-based optimizer extend naturally to discrete or combinatorial latent structure, or is it tied to continuous Langevin dynamics?

Your PPDL paper treats LLM flows as probabilistic programs with SMC-based inference - that's almost exactly the compilation pipeline we've been building from ChiRho/Pyro into our world-model agents at Basis; how are you handling structured control flow (branching agent steps) ins

Your line from RankMe to V-JEPA to latent-action world models reads like a bet that representation quality and world-model quality are the same problem — curious how far that extends to non-visual, symbolic world-model settings like MARA's.

You and Kyurae Kim have built a rigorous convergence theory for black-box VI - Basis hits the MCMC-vs-VI tradeoff constantly scaling probabilistic world models; curious whether your linear-convergence guarantees say anything about when full MCMC beats amortized VI in a MARA-style

brms basically made hierarchical Bayesian modeling a one-line affair for a generation of applied scientists - the open question for us is how much of that intuitive-prior, hierarchical-modeling philosophy transfers when the model is amortized and doing simulation-based inference

BayesFlow turns simulation-based inference into something you can call in real time rather than run offline for hours - that's the missing piece for Basis agents that need to update a causal world-model on the fly as they run new experiments; where does amortization break down wh

Your location-scale noise model work for causal discovery fits how we think about world-model discovery under real, non-Gaussian sensor noise — how would the skewness-robust identifiability result change if the 'noise' is itself the byproduct of an agent's own exploratory actions

Your compression-progress account of curiosity is basically the theoretical spine under Basis's active-experimentation approach to world models — I'd love to hear how you'd frame MARA-style embodied abstraction-discovery in that same interestingness/compressibility language.

Causal-JEPA's object-level latent interventions and your LeJEPA work on ditching the heuristics both push toward the same thing MARA is after -- world models an agent can actually intervene on, not just predict with. Since you're working with Heejeong Nam, who's also on our radar

Decrypto is a deception/communication game stress-testing theory-of-mind in LLM agents -- close to how MARA thinks about active experimentation, where an agent models another agent or the environment through interaction. Curious if LOLA-style opponent-shaping is compatible with s

Your aircraft-design paper turns diffusion models loose as a design-space explorer under discrete+continuous SBI constraints -- structurally the same propose/simulate/update loop MARA is building for robot experimentation. Curious how that formulation would generalize to closed-l

Your local-BayesOpt convergence work treats 'what to try next' as local optimization — MARA asks nearly the same question for choosing physical experiments with RoboMARA; do local BO's guarantees survive an embodied, non-stationary action space?

Your discriminative-calibration and predictive-VI work is exactly the workflow-level diagnostic Basis needs when agents synthesize and compare many candidate probabilistic programs on the fly rather than fitting one fixed model — how would you adapt predictive-optimality to a mod

We care about inference staying robust when the simulator is wrong, not just when it's slow — how far does your misspecification-robust SBI work generalize to settings where the model class itself is actively being revised by the experimenter, closer to what we're chasing with wo

Your Semi-Modular Inference framing of how much to trust each part of a misspecified model is a problem Basis hits when composing probabilistic programs from partially-validated components — curious if the new amortized NPE approach could plug directly into a PPL's inference back

Your ICML 2024 position paper argued Bayesian deep learning matters more as models scale, not less -- the same bet Basis makes building probabilistic programming into MARA's active-experimentation loop; curious whether your mean-field variance-overestimation result changes how yo

Your relational SCM extension is close to a problem we keep hitting in MARA -- representing multi-object, multi-agent causal structure in world models built through active experimentation. Do you see a natural bridge from the relational/combinatorial identifiability results to se

Your ICML paper on minimum-distance summaries for robust NPE tackles exactly the misspecification failure mode we worry about pushing amortized SBI onto real simulators — curious how far the robustness guarantees carry once the simulator itself is a differentiable world model rat

Your modified EAT estimator basically buys back the moment-matching guarantee that made people nervous about gradient estimators for Poisson latents in the first place -- have you tried it inside anything closer to a full sequential/state-space inference loop, or is the POGLM spi

Your Nesting Particle Filters paper at ICML this year is basically sequential Bayesian experimental design for dynamical systems - that's close to the closed-loop world-model experimentation Basis needs for MARA; curious whether nested SMC scales to high-dimensional action spaces

Edward was one of the first TensorFlow-based PPLs and sits in the lineage our ChiRho/Pyro-style stack grew out of - I'd love your read on whether entropic-regularized mean-field VI (per your paper with Bohan) could plug into higher-dimensional world-model inference, the problem B

Your Bayesian-optimisation-on-graphs and automatic-quantum-device-tuning work is the kind of 'turn expensive real-world experimentation into an efficient active-learning loop' problem we're chasing with MARA's robots - though curious how much of your day-to-day now is Mind Foundr

LiNGAM is basically the ancestor of the causal-discovery work we build on at Basis - I'd be curious what you make of your student's move to drop the non-Gaussianity assumption in the multi-view ICML 2026 paper, and whether that generalizes to the kind of active, multi-environment

Tigramite treats causal discovery over time as a first-class systems problem, not just an algorithm - that's exactly the muscle Basis's world-model agents need when they infer causal structure from their own actively-collected interaction data rather than a fixed observational ti

Your CELLO and 'Causal Evaluation of LLMs' work is asking exactly the question Basis's causal-agent stack needs answered — can a model's causal graph reasoning actually be trusted — and it maps directly onto how we're trying to get ChiRho-style causal world models to hold up unde

Causal-JEPA's object-level masking for inducing causal structure is very close to how we think about active-experimentation world models at Basis — I'm curious whether the counterfactual VQA gains held up when you tried interventions the model hadn't seen paired object relations

RT-2 and Do-As-I-Can treat the robot's world model as something you distill from web-scale priors plus affordance grounding — we're approaching MARA's robot stack from the causal/active-experimentation side instead, where the agent has to build its world model by intervening, not

RT-1/RT-2 and Physical Intelligence's whole bet is that scale plus web-scale priors gets you generalizable robot control — MARA is approaching the same real-world-robot problem from the other direction, building a world model through active experimentation rather than imitation-a

Your PriorCVAE/PriorVAE line is doing exactly the kind of thing we want ChiRho and our SBI stack to do — replacing expensive MCMC with a learned generative prior that still gives calibrated uncertainty; how are you thinking about extending that flow-matching correction to structu

Your Dreaming-in-Code paper has the LLM writing the executable environments that the curriculum then trains against - that's close to the loop we want MARA agents to run on themselves at Basis; how far can you push the LLM's role from generating environments to generating and rev

PACER's trick of guaranteeing acyclicity by construction rather than penalizing it is exactly the kind of causal-discovery building block we'd want composable inside a larger world-model — have you thought about what it would take to run PACER online, inside an agent that's choos

Multi-task Bayesian ICL is basically the hierarchical-prior version of the amortized inference problem we work on for PPL-style world models -- does your multi-task setup let the model infer which task-level prior it's in, or is task identity assumed known at inference time?

Your BoTorch/Ax stack has become the de facto substrate for a lot of applied Bayesian optimization -- I'd be curious how you're thinking about LILO's use of natural-language feedback as an acquisition signal, since that's close to how we think about steering active-experimentatio

LIVE's approach to long-horizon interactive video world modeling and your DriveWAM work on action-conditioned generative priors both tackle the same core problem MARA is chasing -- a world model an agent can actually plan and intervene with, not just predict from. Curious how you

Your MineDojo/Voyager arc treats an open-ended world as a self-bootstrapping curriculum -- DreamDojo extends that to real human video; MARA asks the same question via active experimentation rather than passive video. Curious how you'd close that active/passive gap, and whether pr

Score identity distillation is one of the cleaner one-step diffusion-distillation tricks out there, and your SPIGM talk on structured probabilistic inference sits right next to what we're doing with ChiRho's effect-handler-based generative models. Given you're already splitting t

Concordia's framing of generative agents acting in physical, social, and digital space — and 'machine culture' more broadly — is close to the problem MARA runs into: agents that need to build and update a model not just of the world, but of what other agents in it believe and wan

Your empowerment/control-as-supervision framing (DIAYN through the 2026 control-maximization unification work) is close to how we think about active experimentation in MARA -- curious how you'd frame empowerment-driven exploration for an agent that has to build its own world mode

BALD and BatchBALD are basically the acquisition-function backbone for 'what experiment should I run next' — which is the exact question MARA's agents face when they have to design their own interventions rather than just query a fixed dataset. Do you think epistemic-uncertainty-

Your conformal-alignment and 'act-or-defer' work on knowing when to trust a foundation model's output is exactly the certification problem our agents hit once they start proposing their own experiments in MARA -- would love to hear how you'd think about calibrating trust for an a

Your VJEPA framing of joint-embedding predictive architectures as explicitly probabilistic world models is close to how we think about MARA's world-model layer -- I'm curious whether you see a path from JEPA's implicit latent representations toward the more structured, compositio

Your local-learning-coefficient work on stagewise development in transformers is basically asking 'what does the model believe it's uncertain about, and when does that change' — that's the same question we're chasing with world-model uncertainty in MARA. Have you tried applying t

SGLRW's trick of pushing stochastic noise only into the off-diagonal update covariance is exactly the kind of robustness we'd want under the hood of a ChiRho probabilistic program when the data gets messy - with Normal Computing as co-authors, did the thermodynamic/physics-based

You and Michael were the two people who had to actually get the modified EAT estimator working end-to-end on the spike-train GLM experiments -- what part of the moment-matching fix gave you the most trouble in practice?

Your damage-adaptive MAP-Elites work is basically the founding case study for 'a robot that discovers its own model of itself through experimentation' — that's very close to what we're trying to get MARA to do generally, not just for locomotion recovery but for open-ended scienti

Causation, Prediction, and Search is still the book we point people to when they ask what a causal world model even has to assume to be identifiable — I'd be curious what you make of running PC/FCI-style discovery inside an agent that's actively choosing its own interventions rat

The martingale/coherence framing in 'From Drift to Coherence'—checking whether an LLM's sequential predictive beliefs behave like a well-calibrated Bayesian updater rather than drifting incoherently—is close to a question we ask about our own agents' belief states at Basis during

ModelSMC's treatment of LLM-proposed scientific models as SMC particles keeps a principled posterior over hypotheses instead of taking the LLM's top-1 guess — curious how sensitive discovered-model quality is to proposal diversity versus the SMC weighting itself.

FUSE's Feynman-Kac-steered sampling is a neat way to correct amortized posteriors with likelihood information at inference time without giving up the speed of flow matching - have you tried steering with a learned (rather than the true) likelihood when the simulator itself is exp

Decrypto's framing of theory-of-mind as a measurable multi-agent reasoning benchmark is close to problems we care about in MARA -- curious whether the regret-approximation ideas from your curriculum-discovery work could transfer to curating harder ToM curricula, and what you're f

Your Stein Diffusion Guidance paper frames posterior correction as a stochastic-optimal-control problem rather than the usual heuristic guidance-scale tricks -- that's the kind of principled sampling correction we care about for calibrated inference in world models. Have you trie

Your structured 4D latent model splits the representation into explicit 3D scene structure rather than a flat latent - the same bet we're making with the Droid-stack world models at Basis. How well does that structure hold up when the robot's own actions break the rigid-body/stat

Your amortized-VI work for partial-label learning is solving the same 'infer fast under structural uncertainty' problem Basis runs into constantly in probabilistic programming — curious whether your distributional-regression machinery could give calibrated uncertainty over world-

Your LAC-WM cross-embodiment world model and the VJEPA2 action-conditioned planner are basically the industrial-scale version of the world-model + active-experimentation loop we're building for MARA's robots -- I'd love to compare notes on where latent-action representations brea

Your SMC-based stochastic optimization work and the Jarzynski-adjusted Langevin algorithm for latent-variable EM both attack the same particle-based inference bottlenecks we hit scaling probabilistic programs in ChiRho -- curious how your SMC optimizer handles high-dimensional la

WestWorld's knowledge-encoded trajectory world model for diverse robot morphologies is tackling almost exactly the generalization problem MARA runs into scaling world models across our robot fleet -- how are you handling zero-shot transfer to genuinely novel embodiments versus in

Your tractable EIG result for exponential-family posteriors is the kind of closed-form trick that could make expected-information-gain-driven active experimentation actually tractable at the scale MARA needs for real-world robot experiment design -- how far do you think the expon

Your amortized generalized-Bayes estimator learns the beta-conditioning directly, skipping per-dataset MCMC entirely - we're hitting a similar wall doing SBI over MARA's simulated-experiment world models; does the single-network amortization hold up when the simulator itself is n

You've worked the full stack of this question—from anatomically-constrained CNNs explaining retina-to-cortex representations, to Barlow Twins-style self-supervised learning, to now mental rotation as a testbed for world models—which is close to how we think about MARA's active-ex

The on-the-fly prior adaptation in Distribution Transformers is basically the amortized-inference problem we keep running into scaling ChiRho to new causal model classes — how well does that GMM prior-to-posterior mapping hold up when the 'prior' is itself the output of a structu

You've just wrapped the Genentech chapter and mentioned wanting to broaden the aperture on AI conversations -- the amortized hierarchical Bayesian in-context learning work you're on is basically the inference-time story we're chasing with ChiRho and MARA's world models; curious w

Your hippocampal-entorhinal world model does structure abstraction the way we're trying to get MARA's agents to abstract task structure from a handful of physical experiments - do you see the entorhinal grid-like codes as a general-purpose abstraction prior, or something narrowly

Your 'Time for Sampling Is Now' position paper is basically the argument we have internally about ChiRho — that scalable posterior sampling, not just VI point estimates, is what real Bayesian deep learning requires. Where do you think that argument breaks down once the models are

MindCraft's theory-of-mind modeling for collaborative dialogue and RoboMME's memory benchmark both touch something we run into constantly in MARA: an agent needs to track not just its own world model but a collaborator's beliefs about it, especially across a long horizon of physi

Your conformal-prediction-under-feedback-covariate-shift work addresses exactly the failure mode we worry about for MARA's active-experimentation loop -- that the agent's own choice of experiments shifts the distribution the calibration was fit on. Does the covariate-shift correc

Your survey on model-based RL and this year's world-model work for vision-language-action models both frame the same problem MARA is chasing on the robotics side -- using a learned world model to make embodied decisions -- I'd be curious how your VLA-MBPO approach compares to usi

Your posterior-sampling RL work with Gaussian processes for continuous control is doing principled Bayesian exploration in exactly the regime Basis's MARA robots need — active experimentation under real uncertainty rather than just reward-maximizing rollouts; would love to compar

You built EasyRCA to find root causes in live IT-monitoring time series at EasyVista -- that's structurally the same diagnostic problem MARA's robots face when a physical system misbehaves. How much of your JAIR survey's identifiability theory survives when you swap server logs f

Your push on 'the science of scaling' and epiplexity for data selection is close to the theoretical grounding we want under AutumnBench and MARA's world-model evaluations -- curious how generalization bounds hold once reasoning is done by a program-synthesis loop rather than a si

Your ICML paper shows transformers implementing GD toward the GP posterior-predictive mean/variance in-context - close to the amortized-inference story in Basis's Pyro/NumPyro-descended stack; have you thought about extending the binned posterior-predictive output beyond GP regre

Minimum-distance summary statistics for robust NPE hits the misspecification-robustness problem we run into pushing SBI onto messy real-world simulators at Basis - how does your approach handle summary statistics that are only weakly identified?

Your work connecting natural-gradient VI, continual learning, and now federated ADMM under one 'knowledge-adaptation' lens is close to the unifying probabilistic-programming story Basis wants — how far does that lens extend to online model discovery in an active-experimentation l

Your work on identifiability of latent hierarchical causal structures — the tensor-rank conditions for discrete latent variables in particular — is the kind of guarantee we need before trusting a causal world model with unobserved confounders; how far do those results extend once

Your Pandora's Box Gittins-index framework decides when to stop paying for more Bayesian-optimization evaluations — that's exactly the question a MARA robot faces deciding whether to keep probing an environment or commit to acting on its current world model. Have you thought abou

Your and the AMLab group's work threading equivariance and Bayesian deep learning through to BLIPs tracks closely with the inference machinery underlying Basis's probabilistic-programming stack (Pyro's lineage) — curious how you see generalized/equivariant priors interacting with

LABO's gating criterion for switching between LLM-predicted and real experimental evaluations is basically a bet on when to trust the model's world-view over ground truth -- the same question MARA agents face deciding whether to simulate or physically probe an environment.

MCLMC's scaling story -- 4 hours vs. 3+ days against NUTS at 2M parameters -- is the kind of inference speedup that would let MARA's world models condition on far richer interaction histories; is the Metropolis-free tuning scheme robust enough to drop into a probabilistic-program

Your RED-HDP-HMM makes state durations observation-dependent instead of stationary — we've hit similar expressiveness limits building explicit-duration/effect-handler style composition into ChiRho; how would you feel about implementing RED-HDP-HMM's Gibbs sampler as a reusable pr