Language models can retain the numbers or fields an answer requires while putting them at the wrong structural address in a JSON output or table, according to an arXiv preprint. The authors hypothesize that this gap between having a value and placing it correctly becomes more prominent as structures grow more complex.
The error inside a pass-fail score
To make that gap visible, the paper uses Structure-Content Decomposition, or SCD. The method separates three questions: whether the output is valid in its format, whether it follows the requested schema, and whether each value sits in the required position. That lets an evaluator distinguish a malformed answer from one that contains the right value but attaches it to the wrong location.
The diagnostic comparison covered six models spanning open-weight and closed-source systems. Each model faced 1,500 JSON tasks, with 500 at each of three complexity levels, and 600 table tasks, with 200 at each level. The tasks were generated algorithmically without an LLM. Required values were planted into specified structural positions, so the tests had deterministic ground truth.
Complexity opens the gap
At the highest JSON complexity level, the share of values present was higher than the share placed accurately for the models highlighted. DeepSeek-V4-Flash misplaced 35% of the values it recalled, while Qwen2.5-7B misplaced 74%.
Additional comparisons separated depth from the number of values in a task. In Qwen2.5-7B JSON tests, displacement was near zero at tree depth 1 and 28.5% at depth 3. Across planted-value counts from 5 to 20, the displacement rate showed no monotonic trend, staying around 56% with fluctuations of about 5 percentage points.
Other M-level comparisons reported displacement of 49.6% for opaque navigation, versus 42.1% for descriptive names. Repeated names were associated with 52.8% displacement, compared with 44.6% for unique names.
The same problem appeared in tables
The pattern also appeared in tables, although the models differed sharply. At the highest table complexity, strong models had displacement rates between 17% and 28%, while Qwen2.5-7B reached 98%. For Qwen, value presence was only 0.085. The results keep format validity and coordinate placement as separate measures, rather than treating a recognizable table as a complete success.
A reward aimed at placement
The paper then reports a structure-aware reinforcement-learning test called SA-RLVR. It used GRPO with 10 sampled completions per prompt and trained Qwen2.5-7B-Instruct with LoRA for 500 steps on about 3,400 mixed-domain prompts. The reward assigned value-placement accuracy a weight of 1.0 and schema compliance a weight of 0.3. Value presence was excluded because it could reward a value that appeared somewhere in the output without being in the required position.
On JSON-ID, the SA-RLVR-trained model had a reported value-placement score of 0.629, compared with 0.264 for the base model, a reported relative difference of 138%. Its value-presence score was 0.869, versus 0.310 for Base. A matched supervised-fine-tuning baseline scored 0.281 on placement, close to Base.
The reward comparison put those results in context. The composite reward recorded VPA, VP and SCR scores of 0.629, 0.869 and 0.832. A VPA-only reward recorded 0.621, 0.696 and 0.625, while an exact-match reward recorded 0.579, 0.831 and 0.789. Among the tested variants, the composite had the highest reported combination of placement, value presence and schema compliance.
Table evaluations were less encouraging on coordinate placement. On Table-ID, SA-RLVR's format-validity score was 0.855, compared with 0.265 for Base and 0.270 for SFT. Its placement score was 0.065, versus 0.027 and 0.032. On Table-OOD, format validity was 0.940, but placement was 0.085, compared with 0.094 for Base and 0.086 for SFT. In these comparisons, the clearer transfer was in format validity, not table coordinate placement.
What the test can show
The study's evidence is tightly bounded by its design. The diagnostic cases were controlled synthetic outputs with planted values and predetermined target positions, and the training test used one 7B model with LoRA, about 3,400 prompts and 500 steps. The table results also showed little placement gain on out-of-distribution tasks.
Within that scope, the paper makes a practical case for splitting structured-output scores into separate questions: did the required value appear, and did it appear at the required structural address? The document is an arXiv preprint, version one, dated 26 August 2026.
Paper data and sources
Original title: Where vs What: Decomposing Structural and Content Failures in LLM-Generated Structured Outputs
Authors: Yiwei Zhang, Chengke Wu, Li Wang, Jianqiang Li
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text