Research
The question running through my research is: how should models learn from their own generated computation (rollouts, search traces, latent futures) while preserving distributional coverage and knowing when their internal representation is sufficient for the decision at hand? Underneath it is a formulation question. The choice of training formulation shapes what a model can represent, where it fails, and how detectable those failures are. I think of this as representational/model competence under distribution shift.
The common object across these settings is a single loop: acquire feedback through search and exploration (deciding where to query), assess its quality by critique and filtering, then amortize it back into the base model.
- Search creates supervision. Generated computation (rollouts, verifiers, self-consistency, process supervision, latent futures) produces structure worth internalizing.
- Distribution-aware correction, or OPD, internalizes it without collapse. The model learns on its own induced states while preserving the diversity and coverage of the pretrained or reference distribution.
These questions show up in diffusion distillation, reasoning post-training, and embodied world models. I treat them as different instantiations of the same loop rather than separate topics, connecting my work on diffusion distillation, open-set recognition, efficient deployment, multimodal representation, and embodied control.
Generative AI & Efficient Inference
World’s first on-device generative AI. I was ML lead in the cross-functional Qualcomm team that shipped the world’s first and fastest on-device text-to-image generation: sub-0.6s Stable Diffusion on Snapdragon hardware. I built the training pipeline, managed 50TB data pipelines via MosaicML MDS, worked through progressive distillation and W4A8 quantization constraints, identified the manifold-thresholding technique that resolved diversity collapse, and modified diffusion sampling for on-target quality. Featured at MWC’23 and Snapdragon Summit and covered by The Verge and Engadget.
Diffusion distillation as learning from rollouts. Standard diffusion distillation trains a student on teacher outputs at fixed steps, but at inference the student conditions on its own previous outputs. That creates a trajectory mismatch: errors compound across denoising steps, intermediate marginals drift, and diversity collapses. I treated this as a sequential decision-making problem, where the student is a policy, its denoising path is a rollout, and teacher/student mismatch is the standard covariate-shift failure mode from imitation learning.
The fix is DAgger-style mixed-distribution correction (DDIL): train on student-induced states to correct the distribution the student actually visits, while mixing in teacher and reference states to preserve diversity and coverage. I first applied this to progressive distillation, which helped enable the on-device deployment above, and later extended the framing to on-policy distillation, where preserving intermediate marginal distributions remains the central challenge and concurrent methods (e.g., DMD2) tend to lose diversity.
From efficient sampling to search amortization. The transferable idea in DDIL is not “make diffusion faster.” It is amortizing a slow teacher, search process, or scaffold into a faster model without collapsing the distribution that made the scaffold useful. That is the same structural problem that appears in reasoning post-training, where verifier-filtered samples, process supervision, self-consistency, tree and search traces, and on-policy rollouts are all generated computation that can become supervision.
Efficiency as retention, beyond throughput. Deployment makes efficiency concrete, but the research metric I care about is retention: how much useful trajectory behavior (refinement, self-correction, diversity, recoverability) can survive at a fixed compute budget? Compressing an iterative model changes the states it visits, which is itself a new distribution shift.
RL for continuous diffusion. I’m also interested in how the DDIL and DAgger machinery transfers to reward-aligned continuous generation (diffusion-RL over the denoising trajectory), diffusion policies for continuous-action manipulation, and continuous-diffusion and video world models as substrate for downstream policy and value learning. This is a natural extension of the diffusion work toward the reasoning- and control-side directions below.
LLMs, World Models & Embodied AI
In agentic AI, across both language and embodied systems, the same question becomes operational: when search produces better behavior, how do we turn it into model capability while preserving coverage and knowing when that capability can be trusted?
Post-training. A model samples candidate solutions, a verifier or process signal selects or scores them, and post-training pushes the model toward the induced distribution of successful traces. This creates the same risks as diffusion distillation: exposure bias (covariate shift), verifier overfitting, reward hacking, and coverage collapse.
This is where much of my current attention sits: search amortization and on-policy correction for reasoning, training on self-generated rollouts while preserving the diversity and robustness of the base model, and studying whether the model has internalized useful search rather than merely imitating scaffold artifacts. I’m interested not only in whether a recipe improves reward, but in the mechanism behind it: how distribution-aware correction shifts training-signal variance, teacher/student overlap, feature alignment, credit localization, verifier dependence, and coverage, evaluated against matched controls across the design space of OPD, RL, and the post-training toolkit.
Embodied world models. In embodied AI, the same loop moves from reasoning traces to action and latent-future rollouts. Vision-language-action systems are strong at language-conditioned task planning but are largely open-loop; Where as video-diffusion(WAMs), and latent world models offer richer future-state and continuous-control generative priors, but need mechanisms for correction, monitoring, and replanning. The TAMP framework names the divide: VLAs are model-free, feedforward planners with strong LLM-native agentic reasoning; video and diffusion world models are model-based feedback controllers with physics-aware priors that feedforward reasoning structurally lacks.
Two questions follow. (1) Search: what controllable intermediate bridges these paradigms at scale into a unified substrate across memory, reasoning, and action? (2) Competence: what is the interface, the representation, rollout-correction, and competence-monitoring layer between high-level reasoning and closed-loop control, that applies the same search, correct, estimate-competence, and control/feedback loop of the LLM post-training toolkit to action and latent-future rollouts, deciding when to replan, query a simulator, switch controllers, or defer? Both matter for the same reason: search finds the behavior, but without a competence-aware interface to gate it, neither language nor embodied systems can tell when that behavior is trustworthy. Reasoning post-training remains the faster-feedback place to develop both questions before they have to survive contact with a physical system.
Substrates and controllable abstraction.
Autoregressive, diffusion, recursive, and world models impose different commitment structures on intermediate computation. The useful question is: which representation are we training and correcting over, and does that representation support intervention, refinement, and competence estimation? Work such as Skip-to-Good-Part gives evidence that diffusion and autoregressive models organize intermediate representations differently; I view that as a lever for studying controllable abstraction.
Competence Estimation, Self-Knowledge & Failure Detectability
What a representation preserves bounds what any monitor can know. My graduate work at Oregon State (advisors: Tom Dietterich, Alan Fern) studied open-set recognition and failure detectability under distribution shift. The core lesson: a downstream classifier, confidence head, verifier, or selective-prediction rule cannot recover competence-relevant information the representation has already discarded. The leverage is upstream, in the training objective and representation, not only in a score attached afterward.
Open-set detection and reliable representations. I studied which representation structures (contrastive, ensemble, flow-based, VAE-based) make unknown-class inputs detectable and where they break down. Pairing discriminative objectives with generative priors raised the ceiling on detectability because it preserved more of the data manifold. That principle now carries into reasoning and planning: a process reward model, verifier, or monitor is only as good as the intermediate state it reads from.
Risk-sensitive RL and constrained MDPs. A concrete instance of competence as a gate: incorporating uncertainty into policy learning through constraint-based methods that penalize acting where the world model is unreliable. This is the policy-learning face of knowing when not to trust a rollout.
Online sufficiency. A question I’m interested in is whether competence can be assessed online, from intermediate computation (hidden states, partial reasoning traces, rollout latents, feature trajectories, cross-modal abstractions) early enough to defer, replan, resample, or abstain before the answer or action fails. The same idea instantiates differently across domains: an unsupervised divergence monitor over reasoning-trace hidden states (test-time-compute control); competence-routed depth and compute in cross-modal alignment; supervised multi-source mismatch gating search, safety, and simulator queries in embodied rollouts. The connecting thread is one: competence is upstream of the head, and online sufficiency is the runtime instrument that closes the loop.
Last updated: June 2026