Source-Preserving Prompt Augmentation and Category-Adaptive Inference for Qwen3-1.7B: A Prospective Seed-Replication Study

Research question

Can a fixed compound inference intervention improve the objective performance of a small language model without updating its parameters?

The study examines a systems-level alternative to model scaling: preserve the original request, add source-supported control information, and allocate inference-time computation according to task category.

Experimental design

The prospective P1.3 replication evaluated Qwen3-1.7B on a fixed set of 96 public tasks: 48 from IFBench and 48 from LiveBench, with research-authored semantic-stress variants. It reused the frozen tasks, augmentations, model revision and evaluators from the preceding study while introducing two new stochastic seed roots.

Four conditions separate the two parts of the compound intervention:

• raw prompt with non-thinking inference;
• raw prompt with category-adaptive inference;
• source-preserving augmentation with non-thinking inference; and
• source-preserving augmentation with category-adaptive inference.

Scores were averaged across seeds within each task before statistical inference. This design tests robustness to new stochastic draws; it is not an independent task-level or cross-model replication.

Primary result

Comparison	Mean objective score	Paired effect	95% bootstrap CI	Wins / ties / losses
Raw, non-thinking	0.272	Not applicable	Not applicable	Not applicable
Augmented, adaptive	0.360	+0.088	0.018 to 0.160	24 / 63 / 9

The complete intervention produced a 32% increase relative to the raw non-thinking mean and met the prespecified positive-classification rule. The confidence interval excluded zero, but its lower bound remained below the prespecified +0.05 practical threshold. The result therefore establishes a positive effect under the study’s classification rule. It does not establish that the population effect necessarily exceeds +0.05.

Component evidence

The decomposition is central to the causal interpretation.

Component contrast	Effect	95% bootstrap CI	Interpretation
Adaptive inference profile on the raw prompt	+0.061	0.021 to 0.107	Supported
Adaptive profile within augmentation	+0.043	0.002 to 0.092	Positive; below the frozen +0.05 practical threshold
Augmentation under matched adaptive inference	+0.027	-0.046 to 0.099	Directionally positive, but uncertain

The experiment confirms the complete package relative to raw non-thinking inference. It does not establish semantic augmentation as the active ingredient or a beneficial interaction between augmentation and the adaptive profile.

Robustness checks

Both seed-specific point estimates were positive: +0.108 for seed 20260829 (95% CI: 0.026 to 0.190) and +0.068 for seed 20260830 (95% CI: -0.009 to 0.147). The second seed was individually inconclusive, but the task-averaged interval excluded zero.

Prespecified sensitivity analyses preserved the conclusion:

• excluding three tasks with unresolved augmentation artifacts produced +0.093 (95% CI: 0.021 to 0.168; 93 tasks); and
• contract-aware secondary scoring produced +0.085 (95% CI: 0.016 to 0.156).

LiveBench showed a +0.103 effect (95% CI: 0.019 to 0.195), while the +0.073 IFBench estimate crossed zero. These benchmark and stress-family analyses were secondary, used small subgroups and were not adjusted for multiplicity.

Systems cost

The quality gain required materially more test-time computation.

Condition	Input tokens	Thinking tokens	Final / total output	Mean latency	Total GPU time
Raw, non-thinking	465.17	0	842.31 / 843.14	13.59 s	2,608.80 s
Augmented, adaptive	745.91	1,184.41	718.63 / 1,904.80	30.76 s	5,720.50 s

Peak VRAM remained essentially unchanged because the downstream model was the same. Upstream semantic compilation was a separate one-time cost: the preceding phase processed 208 prompts in 341 provider attempts, accepted 201 artifacts, left seven unresolved and incurred $0.7099 in measured API cost. P1.3 reused those frozen artifacts and made no provider calls.

The paper therefore does not claim a quality-cost Pareto improvement.

Interpretation

The results support treating semantic analysis as a control plane rather than a replacement for task data. The original request remains authoritative; added guidance highlights constraints, precedence and output requirements. A second control layer allocates thinking mode, sampling settings and output allowance by task category.

The practical direction is selective compound inference: bypass simple requests, preserve and augment constraint-heavy inputs, and enable the expensive profile only when the expected benefit justifies its cost.

Limitations

• The confirmatory result is limited to Qwen3-1.7B and one fixed 96-task population.
• New stochastic draws strengthen internal robustness but do not test new tasks, compiler outputs, researchers, implementations or model families.
• The primary comparison changes both source-supported guidance and the inference profile, so it does not isolate semantic augmentation as the causal mechanism.
• The semantic compiler was a proprietary API model without a distinct dated provider snapshot; future provider drift cannot be eliminated.
• Secondary and subgroup intervals were not adjusted for multiplicity.
• The gain required substantially greater latency and GPU time.

Materials and next step

Frozen prompts, source-preserving augmentation objects, raw generations, evaluator traces and analysis scripts are retained in versioned research archives. They are available from the author to qualified researchers upon reasonable request, subject to benchmark licenses and provider terms.

The next scientific step is a fresh-task, cross-model replication with controls that isolate augmentation from decoding profile and inference-time compute.