- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper asks why linear probes calibrated on a world model's encoder outputs fail when reused on latents the model produces with its own dynamics: is the physical information lost, or has the probe's calibration changed? Holding a 5M TD-MPC2 model, its states, actions and edits fixed, the authors compare an encoder-calibrated ridge readout with a readout refit on rolled latents, a target-agnostic linear realignment map from rolled to encoded latents, and no-dynamics baselines, on three DeepMind Control tasks, three state distributions and horizons up to ten model steps. On perturbed cheetah states at k = 5 the encoder-calibrated readout falls from R² = 0.910 on encoded latents to 0.038 on rolled latents, while the refit readout and the realignment reach 0.839 and 0.823, which the authors read as a calibration change rather than lost information. They further report that the realignment is condition-specific and better described as identity plus a low-rank correction than as a rotation, that latent cosine cannot certify a readout (supported by a whitened error bound), and that swapping the readout's calibration source changes intervention-consistency (IC) effect sizes on cartpole by 1.35× in a pre-registered comparison (up to 2.6× across four comparisons) without reversing any feature-versus-control separation. A second architecture, LeWorldModel on Reacher, shows no mismatch on its training distribution.
Critical strengths
-
The experimental practice is careful and unusually transparent. The IC comparison is paired and pre-registered, failed predictions and deviations are reported openly, and a separately pre-registered within-calibration null on disjoint training halves shows that the IC changes exceed refitting noise. All numbers are generated from released artifacts, and values spot-checked against the supplementary result files match the paper. This makes the individual measurements credible, even where I disagree with their interpretation.
-
The diagnostic decomposition is simple and practically useful. Comparing the encoder-calibrated readout against a refit readout, a target-agnostic realignment and no-dynamics baselines cleanly separates "information lost" from "calibration lost". The paper also shows that a latent cosine of 0.947, above the common 0.9 gate, can coexist with a readout R² of 0.038, while cosine still ranks conditions by transfer gap (Spearman −0.808), which gives practitioners a cheap triage tool.
Critical weaknesses
-
Limited novelty, by the paper's own account. Probe failure on rolled latents is prior work (Vakalis, 2026; Guei et al., 2024), and Appendix J claims no novelty for repairing a drifted probe with a label-free low-rank map (cf. Dingeto, 2026; Duan, 2026). Proposition 1 is a loose Cauchy–Schwarz bound (median realised-to-bound ratio 3.2×10⁻⁴ to 0.016), and the whitened-versus-raw point was already made for language-model probes (Ying et al., 2026). The remaining novelty, condition dependence and the IC effect, is weakly supported (W2, W3).
-
The IC evidence is thin. It comes from cartpole alone, the task with the smallest transfer gap, and only one of three pre-registered contrasts has a ratio interval excluding 1 (1.35 [1.02, 1.89]), a bound the authors note would not survive multiplicity correction. The 2.6× headline is a non-pre-registered replication scored on 3 of 5 variables. IC is inherited from an uncited analysis, its magnitude metric is at floor, and nothing shows which calibration is correct on edited latents.
-
The condition-dependence and "not a rotation" claims are weaker than presented. The paper's own C_roll, fit once on rolled latents pooled over k = 1–10, reaches R² 0.86/0.85/0.85/0.79 at k = 1/3/5/10 on perturbed cheetah (seed 1), close to per-horizon refits (0.91/0.87/0.85/0.81), so one calibration appears to serve all horizons; pooling across state distributions is never tried. The symmetric share is computed on W rather than W − I, and for an orthogonal Q it equals (1 + tr(Q²)/d)/2, which tends to 100% near the identity, so 61–70% does not argue against a rotation. SimNorm's 64 null directions force zero singular values in any ridge map, so only the large singular values (up to 17.4) remain as structural evidence.
-
Narrow scope; the large effect sits in a constructed off-distribution setting. All main results use one 5M TD-MPC2 on three proprioceptive DMC tasks with 2–3 checkpoints each. Under the planner the gap is ≤ 0.021 up to k = 5 (0.03–0.14 at k = 10), versus 0.65–0.90 on perturbed cheetah, where the torso is drawn over the full circle under noisy or random actions. The LeWorldModel check is a single checkpoint tested in-distribution only.
-
The presentation obscures the contribution. The text is dense and heavily hedged, and key methodology (IC, gates, map families) sits in appendices or an uncited pipeline. The OpenReview abstract reports only the 2.6× figure, not the pre-registered 1.35×.
Questions for authors
See weakness part.