I work on action-conditioned world models and partially observed agents, and on evaluations that distinguish model capability from artifacts of the measurement. Longer term, I'm interested in how useful behavior emerges through interaction between agents.
At Lossfunk, I develop models on frozen vision encoders that compute camera motion analytically and learn the residual scene dynamics. My work combines experiment implementation, matched baselines, preregistered hypotheses, and checking claims against saved results.
Papers and ongoing experiments. Statuses refer to the work below; the PDFs retain their original manuscript metadata.
World models · 2026 manuscript
Separating Command Error, Chaining Cost and One-Call Error in the Rollout Evaluation of a Video World Model
A rollout's error can grow because the model is being given the wrong commands, because predictions are chained through previous outputs, or because predicting farther into the future is already difficult in one call. A single error curve does not separate those effects.
This work audits released navigation world models from 49M to 1.01B parameters using corrected commands, matched single-call predictions, and repeated-command controls. These comparisons require no retraining.
Result
Correcting the rollout commands lowers the 1B model's 4 FPS LPIPS by 18.8% (95% CI: 16.9–20.6%) on the prospectively specified confirmation cohort.
Method
Paired references on the same items and horizons; decision rules registered before evaluation; saved results checked against manuscript claims.
Scope
One released model family. The split between chaining cost and single-call error is a post-hoc analysis and differs between cohorts.
I develop action-conditioned predictors on frozen vision encoders, using analytic camera transport as a base prediction and learning the residual. The comparison keeps the learned architecture, parameter count, and initialization matched.
Result
In the held-out ARKitScenes DINOv3 experiment, prediction skill rises from 0.393 to 0.557 with transport as the base prediction. Transport alone reaches 0.482; providing it as an input reaches 0.427.
Measure
Skill is one minus prediction MSE divided by copy-last MSE. This result uses a 3M-parameter predictor, two seeds, and 100 test scans.
Contribution
Analytic transport and learned residual models, parameter-matched comparisons, and evaluation with held-out splits and visit-level bootstrap intervals.
Readout is not generation: SE(3) in the patch tokens of frozen encoders
When a camera moves, some of the change in a vision model's representation follows directly from geometry. I study how much of that change can be computed by resampling the patch-token field, with no learned motion predictor.
The experiments compare geometric transport with learned predictors, examine transfer to new scenes and sensitivity to pose errors, and show how mean pooling can hide spatial structure that remains useful for prediction.
Result
A rotation warp with no fitted parameters explains 16–37% of held-out feature change at a four-frame gap and beats the matched learned flow head in 10 of 15 comparisons. The fixed-depth plane warp wins 15 of 15.
Scope
Five encoders in the flow-head comparison, a small-data regime, and cells above a prespecified displacement floor. The plane warp's depth is fixed at 2 m, not tuned on these comparisons.
Self-supervised learning · NeurReps 2026 workshop work · revised manuscript
Reading Relatedness and Hierarchy from Frozen Representations
Which relationships can we read from frozen representations, and at what resolution? This project compares pairwise relatedness, local quartet agreement, and exact reference splits across text, speech, proteins, vision, and single-cell data.
Reconstructed trees recover a minority of the specified splits in the four larger populations, yet retain substantial quartet agreement. Low exact split recovery therefore does not establish that hierarchical information is absent.
Method
Matched readout comparisons, neighbor joining and FastME reconstruction, exact additive-distance controls, and a separate feature-space simulation.
Finding
Changing the origin can substantially change cosine-based associations. Centering largely explains an apparent alignment effect in the language readouts.
Scope
Each domain has its own reference: a semantic taxonomy or cell-type grouping does not describe evolutionary descent. Results concern the specified readouts and reference relationships.
Model behavior & measurement · manuscript under review
What Transfers through Synthetic Code? Separating Security, Style, and Provenance
A teacher model generates code; other models learn from it. If the students inherit both output style and security behavior, which measurements establish a behavioral change, and which simply identify their common origin?
This study compares human annotation, automated security scorers, corpus statistics, and activation probes. The current artifact audit examines output censoring, generation length, and shared model lineage as explanations that the measurements must distinguish.
Finding
Fresh human annotation of prospectively generated 1,024-token responses finds 39/195 unsafe treatment outputs versus 3/78 controls; the automated scorers disagree on the same outputs. Masking responses at token 320 also changes the measured contrast.
Scope
One teacher/carrier per arm limits upstream replication. The human endpoint supports a behavioral contrast in this pipeline; scorer disagreement, output style and shared lineage limit what it establishes about a portable mechanism.
Spectral Perturbation Bounds for Experience Replay: A Bias–Variance Decomposition for Offline Decision-Making
More diverse offline data can improve statistical efficiency while increasing the mismatch between the policies that collected it. This paper studies that trade-off through mixture Markov kernels and the environment's mixing properties, with a theoretical decomposition and synthetic-MDP experiments.
Reinforcement-learning diagnostics · research manuscript
The Bellman Residual Is Not the Improvement Signal
Does a smaller value-function error tell us that the next policy update will help? CertGap investigates the gap between fitting a critic and improving a policy, combining theoretical analysis with logged updates from PPO, SAC, and TRPO.
The project compares the Bellman residual with diagnostics based on start-state critic bias. These are measurements of policy updates; establishing a useful diagnostic is a separate question from building a better policy-improvement algorithm.
I built a recurrent PPO agent for Gen 9 Random Battles with opponent-role belief updates and damage-aware state features. Training mixes checkpoint opponents and random opponents with win/loss rewards; recurrent state and explicit beliefs address information missing from a single observation.
The technical write-up covers the observation design, belief updates, reward choices, and training setup.
Delegation, Ownership, and the Boundaries of the Self in AI-Assisted Programming
What happens when a tool begins producing work that its user did not specify in detail? This paper distinguishes delegation from familiar forms of tool use and examines how felt capability, knowledge of one's abilities, and credited authorship can come apart.
It brings that conceptual argument into contact with a study of AI co-authorship attestations in public commits. The attestation data measure a changing convention of credit; the proposed account of ownership remains a question for further study.
I'm looking for research and research engineering opportunities in world models, reinforcement learning, and multi-agent learning. For experiment details, collaborations, or an unusually good objection, email me or read my CV. The less formal versions of these interests live in the essays.
