Local Mapping Without Iterative Closure: Zero-Shot and Input-Output-Only In-Context Graph-XOR Capability Boundaries in Qwen3

Abstract

Local Mapping Without Iterative Closure Zero-Shot and Input-Output-Only In-Context Graph-XOR Capability Boundaries in Qwen3 DOI: https://doi.org/10.5281/zenodo.21992852 Frozen-protocol pilot study · Not peer reviewed We evaluated zero-shot and input-output-only in-context graph-XOR capabilitiesin Qwen3 models from 0.6B to 8B under prospectively frozen, bounded protocols. In zero-shotofficial non-thinking mode, reliable behavior reached a one-edge relation at8B but not a two-input XOR behavioral pass. Under the jointcorrect-versus-label-shuffled formation criterion, the two-input mappingsurface first qualified at 16 demonstrations in Qwen3-4B and at 4demonstrations in Qwen3-8B. Correct-demonstration cells passed behaviorally at16 and 64 shots in 4B and at 4, 16, and 64 shots in 8B; only 16 and 64 in 4Band 4 in 8B satisfied the joint formation contrast. No correct-demonstrationlength-8 parity cell met the predeclared score-signal criterion on the frozenledgers. The observed boundary therefore lies between context-conditioned local mappingacquisition and iterative composition. In the ICL branches, shuffledrelational-cycle evaluation remained unopened after parity failed to qualify;the zero-shot scout separately measured ordered and shuffled relational-cyclesurfaces and found no detected signal. Multi-cycle global obstruction remainedunevaluated. The result does not establish areusable XOR algorithm, a specific internal state-tracking mechanism, orabsence of hidden information. A2-M, decorated-theta B, activation extraction,naturality, and causal intervention remained unopened by conditional stopping.Each ICL branch used one prospectively fixed demonstration bank and order;bank-to-bank and order-to-order variance were not assessed. The record contains one integrated 7-page technical report, a publication-readycapability matrix, and one reproducibility capsule with public-safe plans,runners, exact result summaries, cell-level metrics, model/runtime bindings,and checksums. It is a standalone graph-XOR capability-boundary report, not anew version of the author's replay/artifact-structure reports. Suggested keywords graph-XOR; parity; in-context learning; Qwen3; capability boundary; frozenprotocol; negative result; algorithmic reasoning Scope boundary The report closes only the tested Qwen3 0.6B–8B zero-shot officialnon-thinking and bounded input-output-only ICL graph-XOR route. It does nottest thinking mode, generated scratchpads, fine-tuning, models above 8B, otherfamilies, hidden-state information, naturality, causal use, or broaderhypotheses about latent global-obstruction representations.