Decision Model Factory: Building Specialized Sovereign Decision Models

Yuma Ichikawa

cs.AIOct 26, 2026

Decision Model Factory: Building Specialized Sovereign Decision Models

Authors: Yuma Ichikawa

Organizations: Fujitsu Limited · RIKEN Center for AIP

Abstract

Decision needs vary with each application, language, and deployment environment. The Decision Model Factory turns those needs into specialized, sovereign models on demand. A brief recipe guides the training, calibration, verification, and evaluation of a selected backbone, delivering a deployable decision model under the operator's control. The same workflow creates a new specialist whenever the requirements change. Across Japanese decision tasks, compact deployment, and long-context passage ranking, five factory-built models improve upon their backbones; three surpass the strongest results in a fixed public Japanese benchmark comparison. A dedicated runner computes only permitted-answer probabilities and reuses shared context, more than doubling serving speed on matched multi-question workloads with essentially unchanged decision scores. Together, the factory and runner make fast, task-specific decision models practical to build and deploy whenever needed.

Figures & tables

Figure 1. Japanese decision quality on the System One Mosaic Benchmark (S1MB). Red marks show factory-built DF models, hollow circles their backbones before specialization, and black marks public results [ 1 ] .

Figure 2. Decision Model Factory. A model recipe passes through five stages to produce a verified model package and a record of the build.

Model	Gemma 4 backbone	Recipe	Before	After
DF-JA-Large	31B	Japanese-focused	58.99	67.63
DF-Large	31B	General	58.99	66.98
DF-Long	26B-A4B	General, long inputs	51.58	63.57
DF-Small	E4B	General, distilled	31.97	52.96
DF-JA-Small	E4B	Japanese-focused, distilled	31.97	52.49

Table 1. The DF family and S1MB Japanese Task Avg before and after specialization. JA denotes Japanese-focused recipes; DF-Long uses about 4 billion active parameters per token.

Benchmark or test	Metric	DF-Long	Comparison or scope
S1MB English	Task Avg	63.55	DF-Large: 62.31
S1MB Japanese	Task Avg	63.57	Best public snapshot result: 60.05
JQaRA	nDCG@10	90.1	Ruri: 86.9; BGE: 67.8
MIRACL Japanese	nDCG@10	93.7	Ruri: 91.8; BGE: 90.1
MLDR Japanese	nDCG@10	98.9	Ruri: 97.7; BGE: 97.6
Embedded instructions	Ignored (%)	99.9	899 of 900 trials; three languages

Table 2. DF-Long across tasks. S1MB uses 8-bit weights; other results use bfloat16 (BF16) on one NVIDIA B200. nDCG@10 is normalized discounted cumulative gain at rank ten (

\times 100

). Ruri: Ruri v3; BGE: bge-reranker-v2-m3. Calibration uses development data.

Workload	General path	Decision runner	Speedup
Eight questions about one state	71.7 ms	29.4 ms	2.44 $\times$
Eight states, one question each	68.8 ms	30.9 ms	2.23 $\times$
Four states, eight questions each	364 decisions/s	570 decisions/s	1.57 $\times$
One question	16.3 ms	15.0 ms	1.09 $\times$

Table 3. Matched serving comparison for DF-Long in BF16 on one NVIDIA B200. The general-purpose path uses full-vocabulary readout and one request per question. Speedup is the latency ratio, or the throughput ratio for the concurrent workload.

Figure 3. S1MB Task Avg versus eight-question latency. Quality scores use 8-bit model packages; latency uses BF16 and the decision runner on one NVIDIA B200. Timings are separate from Table 3 .

Figure 4. Serving throughput after tuning concurrency and the batched-token budget. Each bar compares tuned throughput with its own default setting; labels give the measured decisions per second.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

	DF-JA-Large	DF-Large	DF-Long	DF-Small	DF-JA-Small
Adapter rank $r$	16	32	32	16	16
Learning rate	$10^{-4}$	$10^{-4}$	$10^{-4}$ / $5\times 10^{-5}$	$10^{-4}$	$10^{-4}$
Effective batch size	32	32	32 / 16	16	16
Maximum prompt (tokens)	8,192	8,192	8,192 / 32,768	8,192	8,192
Teacher weight $\beta$	–	–	–	0.5	0.5

Appendix

Table 4. Training settings for the DF models. A dash indicates that no teacher is used.

LLM Decision-Making

LLM decision-making studies how language models choose, act, seek information, and allocate resources under goals, uncertainty, constraints, and social pressures.

See all 242 papers

this-that-model-1.0: A typed decision model that decides in 30 ms, for a millionth of a cent
Sep 20, 2026Zehua Cheng, Wei Dai, Jiahao SunOn-Device Language Model Inference LLM Decision-Making
University of Oxford
FunDM: Cheap & Label-Free Decision Models from Frozen Language Models
Oct 12, 2026Ujjawal K. PanchalLLM Decision-Making Efficient Language Model Inference
Fundamental AI
Chinese Decision: A Family of Language Models for Single-Pass Probabilistic Decision-Making
Oct 16, 2026Qiang YanLanguage Modeling
Institute of Computing Technology, Chinese Academy of Sciences

Sep 20, 2026J/K move · Enter open · S save

this-that-model-1.0: A typed decision model that decides in 30 ms, for a millionth of a cent

Zehua Cheng, Wei Dai, Jiahao Sun

University of Oxford

On-Device Language Model Inference LLM Decision-Making Efficient Language Model Inference

Open paper AskPDF arXiv

Software delegates more of its branches to models every year: which queue a ticket enters, whether a command is safe to run, whether a claim clears without a person. What the program needs back is not prose. It is one of n declared options and a number it can threshold. Today that costs a round trip to a frontier model -- hundreds of milliseconds, a per-token bill, and a parser -- for a question that is usually a conjunction of three clauses. this-that-model-1.0 is a 2B-parameter typed decision model. Its answer is read directly from the hidden state at a designated position and restricted to the option set the caller declared, so no text is generated, nothing can be malformed, and every question in a request is answered in the same forward pass. It decides in 30.9 ms on one laptop GPU and generates zero output tokens doing it, where a frontier API call costs 8758 ms and the hosted systems that answer these questions well spend between 21 and 212 generated tokens per question thinking first, billed for every one. It sustains 32 decisions per second on one consumer GPU and never lets the state leave the machine. On a third party's recorded cohort of 68 decision questions, on their inputs and their wording, it scores 0.941 with a Brier score of 0.042, against 0.765 and 0.133 for the hosted service Jev on the same items. One pass of our 42-family internal suite takes 32 seconds and 0.000217 USD of electricity; the most accurate hosted model we measured needs 155.2 minutes and 10.636 USD. We also report where it loses. On multi-step arithmetic, which a single forward pass cannot carry intermediate results through, it scores 0.560 against their 0.98 to 1.00, and a targeted second training round improved the five task families it was written for and transferred to none of the other 13. The model is open-sourced in https://huggingface.co/flock-io/this-that-model-1.0

Abstract

LLM Decision-Making

Explore similar work

this-that-model-1.0: A typed decision model that decides in 30 ms, for a millionth of a cent

FunDM: Cheap & Label-Free Decision Models from Frozen Language Models

Chinese Decision: A Family of Language Models for Single-Pass Probabilistic Decision-Making