- Rating
- 1
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper tries to solve the problem occurred in world models trained from serialized logs of asynchronous systems, where the same physical outcome can show up under different action orders. For a state s and two actions a and b, the authors compare the final-state distributions of executing a then b and b then a (the diamond discrepancy), and treat the pair as swappable at s if the difference is below a tolerance. Based on this, the paper proposes three parts. ADD fits a Gaussian process for each action pair over the state space and uses a fixed budget of paired probes to find states where swapping is safe. TRACE-WM then adds reversed-order samples to the world-model training data only where ADD accepts the swap. Graph-TRACE adds a residual model, calibrated with 40 probes, to transfer the two-robot relation to three or four robots. Experiments are on a synthetic intersection benchmark (two robots in the main experiments), with a small MLP that predicts the two-step endpoint as a Gaussian. ADD finds more safe swaps than random probing with the same budget. In the world-model experiments, TRACE-WM has lower prediction error (W2) than the Ordered baseline and a much lower order-effect error than Global invariance.
Critical strengths
-
The problem is well motivated. Logs from multi-agent or distributed systems are serialized, and it is not obvious which order differences are physically meaningful. Learning this as a state-dependent relation, instead of assuming a fixed independence relation as in partial-order reduction, makes sense to me.
-
I also liked that order-effect error is reported separately from average prediction error. Table 1 shows why this matters: Global invariance has lower balanced W2 than TRACE-WM but much worse order-effect error, which would be missed with W2 alone.
-
The paper includes Oracle, Random-derived and FullDyn-derived relations as controls, which makes it possible to check where the gains come from.
Critical weaknesses
-
My main concern is scope. Everything is evaluated on a single synthetic intersection benchmark (2D state and four actions in the main two-robot setting) with a two-step endpoint predictor. All data and the ground truth come from the authors’ own simulator, and the held-out states are taken from the same state pool, so the generalization results are close to interpolation. I also don’t see an easy way to scale this up. The method fits one GP per action pair over a low-dimensional state, and each probe needs to reset to exactly the same state and reuse the same noise. It is unclear how this would work with a latent world model or on a real system. The authors acknowledge some of this, but I think it limits the contribution more than the paper suggests.
-
For the cross-robot pairs, which are the non-trivial part of the benchmark, order only matters when the two robots are close enough to interact. So in this setting, finding where actions commute is the same problem as finding where the robots do not interact. Learning this kind of local independence and using it for data augmentation has been done before (NRI; "Neural Relational Inference for Interacting Systems", or CoDA; "Counterfactual Data Augmentation using Locally Factored Dynamics"), and the paper does not discuss or compare with it. I also suspect part of the problem comes from modeling the logs as single-agent action sequences, since many multi-agent world models just take a joint action at each step.
-
Also, the controls in the paper do not really show that ADD is needed. With a randomly selected relation, TRACE-WM performs about the same as with the ADD relation, and the size-matched control in App. G gives only a small difference. The main gains in the abstract are each against the weakest baseline for that metric. On balanced W2, TRACE-WM with ADD is still worse than Global invariance and the FullDyn relation.
Questions for authors
Q1. How were the training logs generated? For each state and action pair, is only one ordering recorded, or both? And how large is the dataset?
Q2. Could you add an environment where actions interact but still approximately commute, and compare with a CoDA-style relation? In the current benchmark the two notions coincide for the cross-robot pairs, so it is hard to see what the commutativity formulation adds.
Q3. Without resets and shared noise, how would the discrepancy be estimated, and how many samples would it take? I am asking because I could not think of a realistic case where the probes are both possible and expensive. With a simulator like the one in the paper, the discrepancy can be computed for every state (as done for the ground truth), and without one, the paired probes are not available.
Q4. If a learned dynamics model is good enough to read swappability from (as FullDyn does), why not use it directly as the world model? In what setting would TRACE-WM with ADD be the better choice?