Code as Worlds (MirroS technical report, 2026)

30 pages, dated 27 August 2026, authored as "MirroS Team" with no individual named. Read from the PDF. The companion to code-as-world; this is where the method, the numbers and — unusually for a first-party release — two explicit limitations sections live.

Why T3 and not T1. ../QUALITY.md's three-layer rule: the artifact's own facts are T1, the maker's claims about how well it works are T3. Table 1 is a benchmark mirros ran on models mirros trained against baselines mirros configured, and there is no peer review, no venue and no arXiv identifier behind it. The method description would grade T1 on its own; the page is graded by what it mostly carries. Same call already made for unsloth-amd-support.

The argument

A phenomenon–mechanism dichotomy: vision-language models can narrate a physical event but do not hold the object states, physical parameters and governing dynamics that produced it — "describing a phenomenon is not the same as recovering the mechanism that produces it." Recovering the mechanism from partial observation is framed as an inverse problem solved by abductive reasoning — the paper's own word, citing the heliocentric and Newtonian precedents.

The representation is an executable-world-representation; the recovery procedure is an agentic-discovery-loop; the payoff claimed is supervision (§5).

The measured part: QuantiPhy

Eight NVIDIA H100 GPUs, a two-phase curriculum — supervised fine-tuning on image-space measurement, then GRPO on world-level VQA derived from the discovered worlds. 16 frames per video for every model, the official quantiphy protocol, macro-average MRA over the 2S/2D/3S/3D subsets.

The comparison this wiki cares about is the internal one: Qwen3.5-4B scores 31.2 and Code-as-World-VL-4B scores 50.6 on the same benchmark at the same size. That gap is the paper's actual evidence for its supervision, and it does not depend on trusting the proprietary rows.

Two rows are worth flagging for the opposite reason. ChatGPT-5 Pro scores 19.5, below ChatGPT-5's 32.6 and below a 0.26B SmolVLM — a supposedly stronger model far below its own sibling on a benchmark run by a third party with an interest in the ranking. And MiniCPM-V 4.5 scores exactly 0.0 on both 3D subsets, which reads as a harness or parsing failure rather than a capability measurement. Neither is discussed in the report.

The 27B row also changes two things at once — scale and a chain-of-thought response protocol — and the report says so itself: "we view the 27B result as evidence that the framework extends to larger reasoning models, rather than as a controlled estimate of the effect of reasoning alone." The 27B checkpoint is not released.

The loop measured against itself

Video-driven reconstruction, K=5 rounds, compared against one-shot and against Best-of-5 independent sampling at a matched five-evaluation budget — the right control, and rarely offered. Iterative refinement beats Best-of-5 on Visual Alignment, Object IoU, Traj-ADE and Accuracy@2%D. Candidate selection is driven by the loop's own verifier while the reported metrics are held out of the verification signal, so the improvement is not the verifier scoring itself.

Two limitations, stated by the authors

§4.4 — the loop can converge on a wrong world. "When the underlying process falls outside the simulator's modeling scope, the agentic discovery loop may converge to a locally plausible EWR without recovering a mechanistically accurate explanation of the physical process." Verification is against rendered evidence, so a world that renders correctly passes whether or not its mechanism is real.

§5.5 — the model does not learn the loop. "Code-as-World-VL currently learns from supervision derived from verified executable worlds, but does not internalize the agentic discovery process itself… hypothesis construction, simulation, diagnosis, and iterative revision remain external to the model." Also: QuantiPhy covers "only a small part of real-world physics" — no camera motion, rotation, deformation, occlusion, fluids or long-horizon multi-object dynamics.

That second one is the whole cross-wiki finding. See llms-cant-jump, which argues the abductive jump is structurally missing from these models; this report builds the jump as scaffolding around one and then says plainly that the model did not acquire it.

Method, in one pass

• Evidence construction. Text input → entities, spatial relations, events, intended outcomes. Video input → depth maps, instance masks, object tracks, plus a per-object mesh from a 3D generation model (SAM 3D).
• The loop. Algorithm 1: an LLM A and an iteration budget K run propose → instantiate → execute → render → verify, with frame-level discrepancies aggregated into structured feedback that guides the next revision. Terminates on acceptance, or rejects the hypothesis when the budget runs out.
• Supervision. Each verified world exposes exact geometry, camera parameters and a state trajectory, so size, displacement, velocity and acceleration questions can be generated with labels read off the simulator rather than annotated.
• Data. Image-space supervision from RefCOCO/RefCOCO+/RefCOCOg/RefCLEF and GOT-10K; video candidates filtered from WISA-80K by a motion-focused pipeline, then manually reviewed.

Related

• code-as-world · code-as-world-vl · executable-world-representation · agentic-discovery-loop · quantiphy · mirros
• training-data-quality — the opposite strategy for the same problem
• fine-tuning — the SFT-then-GRPO curriculum
• llms-cant-jump (../research-wiki) — abduction as the missing capability