The diagnosis prescribes the fix: keep the goal out of the dynamics and supervise the \emph{read} path, recovering genuine, instruction-independent grounding, and the detection protocol and remedy apply to any goal-conditioned world model whose instruction names the scored quantity.
Abstract
Compact world models that condition on a language goal promise to ground relations such as ``put the red block left of the blue block''using a sparse set of explicit \emph{reference anchors}. We ask when such references actually ground a relation, and identify a trap: a goal-conditioned predictor reaches a striking $0.90$ relation-readout accuracy, yet this is \emph{instruction transcription}, not perception. Withholding the goal collapses it to chance ($0.90\!\to\!0.27$, three seeds) and a counterfactual instruction makes the predicted anchors follow the \emph{false} instruction $94.5\%$ of the time (true scene $2.3\%$; $N{=}256$). Tested across three settings and a within-task ablation, our central claim characterizes the confound: \textbf{instruction leakage occurs when the scored quantity is transcribable from the instruction (when the instruction names the answer) and is essentially independent of how predictive the non-instruction inputs are.} Our tabletop and the external BabyAI benchmark leak, whereas a Language-Table forward-dynamics world model whose instruction names \emph{referents} does not, until the instruction is augmented to name the direction; and degrading the action never increases leakage, the opposite of what predictor-competition predicts. The diagnosis prescribes the fix: keep the goal out of the dynamics (it belongs to the planner's cost) and supervise the \emph{read} path, recovering genuine, instruction-independent grounding ($0.88$, identical with and without the goal). The detection protocol and remedy apply to any goal-conditioned world model whose instruction names the scored quantity.
Humans converge on shared names for novel, hard-to-describe objects through repeated interaction, a process psycholinguists call lexical entrainment. Leading vision-language models fail at this: recent empirical work documents that they do not shorten references, reuse successful expressions, or maintain stable pact state across turns. We present a framework that addresses the gap by externalizing pact state into three explicit, inspectable sets of referent-object bindings ($\Gamma, \Xi, \Omega$), updated by a dynamic-semantics context-change rule. The symbolic layer sits on top of a lightweight perceptual-alignment pipeline that grounds noisy human referring expressions in crowd-sourced imagery via SIFT homographies and the Universal Quality Index. Evaluated on the Stanford Repeated Reference Game corpus (over 15{,}000 director-matcher utterances on abstract tangram stimuli), the framework places the correct target in its top-5 hypothesis set 83.56% of the time from a single director utterance. Human matcher top-1 accuracy on the same corpus is approximately 77-80%. We also report results on a held-out condition in which obvious tangram-adjacent images are excluded from the retrieved set, which provides a more conservative measurement of the grounding signal. Ablations isolate the contribution of each component: SIFT alignment, UQI, query preprocessing, and image augmentation. The central contribution is the combination: a transparent, auditable symbolic layer that recovers the structure of lexical entrainment turn by turn, paired with a perceptual channel whose behavior can be examined ablation by ablation. We also discuss in detail what the framework does not do. It is not interactive, it does not close the loop with the director, and its retrieval-driven perceptual channel is vulnerable to a class of leakage effects that we quantify and bound rather than wave away.
Computer-use agents ground natural-language instructions in screenshots to locate interface elements, yet existing benchmarks do not isolate whether models bind relational language to the correct element. We introduce GUI-Primitives, a 994-item benchmark of contrastive instruction pairs over seven spatial relations in graphical user interfaces (left/right, above/below, containment, alignment, proximity, list ordinal, occlusion). Each pair holds the screenshot and anchor fixed while changing the relation expression, so the correct target moves between two designated candidates. Five annotators validate a 196-item subset ($\kappa = 0.94$ well-formedness; $\kappa = 0.79$ target selection). Nineteen vision-language models reach at most $32\%$ strict point-in-box accuracy. Because models emit unconstrained coordinates, we classify each prediction by the candidate region it falls within. Predictions fall outside both candidates on $60-92\%$ of items. Conditional on falling within a candidate region, target selection reaches 0.82-0.90 for horizontal position, vertical position, proximity, and list ordinal, but does not differ significantly from 0.50 for containment and occlusion: most failures reflect candidate localization rather than relation understanding. Across ten models, benchmark accuracy correlates with ScreenSpot-Pro accuracy (Spearman $\rho = +0.74$), an exploratory association at this sample size. Marking the two designated candidates raises selection accuracy by 35--57 percentage points, an oracle diagnostic that supplies the candidate set rather than a deployable method. We release the benchmark, predictions, and code.
World models built on the RSSM architecture, such as DreamerV3, keep a recurrent hidden state $h_t$ trained only to reduce prediction error. We show this state also tracks its own confusion, hiding in plain sight: nearly orthogonal to $h_t$'s directions of greatest variance, invisible to any variance-based method. It is functionally distinct from ensemble disagreement, which flags new inputs, and reconstruction error, which flags bad predictions right now. On a test holding prediction error fixed while confusion varies, a linear probe on $h_t$ finds the signal (AUROC 0.72, 5 runs), while an ensemble baseline scores below chance. A discounted count of recent high-error steps explains 80% of the probe's output ($R^2=0.80$). We confirm the signal is causally used, not merely present, by editing $h_t$ directly and watching behaviour change, including a check using real values from other trajectories instead of synthetic edits. Its geometry and closed form generalize across three control tasks; the decisive dissociation test itself holds cleanly on only one, and its practical use, deciding when to check reality instead of trusting imagination, generalizes to only two of the three tasks.
We study the language-model head and softmax as a single module, deriving an update geometry from their composition rather than from the weight matrix in isolation. Under Hilbert's projective distance, the maximum change caused by an update $S$ over $\left|\left|{h}\right|\right|_2\le H$ is $H\max_{i<j}\left|\left|{s_i-s_j}\right|\right|_2$, which is $H$ times the Euclidean diameter of its token rows. Motivated by Muon's singular-value conditioning, we propose maximizing the smallest row separation while constraining this diameter, producing an approximate-equidistance problem when $V\gg d$.
We present \textbf{DreamX-Phi 1.0}, an action-conditioned video world model for robotic manipulation that, given an observed frame, a language instruction, and a prescribed action sequence comprising end-effector poses and gripper states, predicts the resulting future observations. Yet realism alone does not guarantee faithfulness: a convincing rollout can still move the wrong arm or lose the manipulated object. To ensure the prediction respects each arm's commanded path, we inject per-arm $\mathrm{SE}(3)$ transformations into attention via \textbf{PRoPE-style geometric encoding}, preserving arm identity and rigid-motion structure. Action control alone does not fully constrain scene geometry or the evolution of small manipulated objects. We therefore add a lightweight \textbf{depth branch} for scene-level geometry and use \textbf{SAM3 masks} with a frozen \textbf{V-JEPA teacher} to maintain object consistency throughout grasping. We further distill the multi-step generator into a few-step student via distribution-matching distillation for efficient deployment. At the time of writing, \model{} achieves first place on Track~1 and second place on Track~2 of the WorldArena~2.0 Challenge. Our model and code will be publicly available.
Dream Team, Rui Chen, Xiangxiang Chu et al.· 1 citation
The results show that referent selection and boundary precision are partially separable, with different components moving opposing regions of the IoU curve -- behavior a single threshold cannot reveal.
Bo Ma· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
