Architecture and reasoning

Generative models, sparse MoE theory, and reasoning controls

The model discussion distinguishes measured extraction behavior from unobserved routing, reasoning, and quantization mechanisms.

Generative extraction

LLM implementation is summarized at the experimental level

ChatGPT, Gemma, and Qwen generate structured text conditioned by extraction prompts. Gemma and Qwen were served through llama.cpp with quantized GGUF checkpoints, 65,536-token context, four parallel slots, Vulkan execution, and reasoning disabled. ChatGPT was recorded as “model 5.6 on high”; provider build and sampling trace were not archived.

Sparse mixture of experts

Gemma and Qwen route tokens through selected experts

yt=iTopK(gt)KptiEi(ht)

The router assigns token representation hₜ to a small top-K subset of experts. Sparse routing changes active compute, capacity, load balance, and quantization sensitivity. Because there is no dense sibling control, router telemetry, or repeated seed design, the observed Gemma–Qwen difference cannot be attributed causally to MoE topology.

Figure 4. Sparse MoE token routingOnly a subset of experts is active for each token; routing behavior was not logged in the benchmark.

Reasoning controls

Reasoning-disabled and reasoning-enabled extraction are distinct treatments

Reasoning-disabled decoding can reduce latency and unsupported elaboration but may weaken reconciliation across distant evidence and schema constraints. Reasoning-enabled decoding may improve planning and self-checking, yet can increase token cost, paraphrase, and unsupported completion. The current local runs used reasoning off; the study therefore cannot estimate a reasoning effect without a factorial rerun.

Non-generative analogue.

HNLP and BERT have no autoregressive reasoning mode. Their corresponding treatment variables are dictionary versions, synonym expansion, optional OpenNLP resources, chunk windows, thresholds, label batches, mapping constraints, and backfill.

Model-selection programme

Alternative architectures to test

Layout-aware encoders

LayoutLM-family systems test whether columns, tables, and spatial headings improve reconstruction.

Dense controls

Dense siblings of Gemma or Qwen isolate sparse-routing effects more credibly than cross-family comparison.

Sequence models

Transformer-CRF or BiLSTM-CRF baselines provide fixed-label sequence constraints and transparent annotation requirements.

Constrained seq2seq

Schema-constrained generation tests long-range synthesis while limiting malformed output.

NextFive-system results →
Figure