Expanding the Pareto frontier of novelty and utility on unmodified foundation models.

For the past three years, post-training alignment created a seemingly rigid law in generative AI: you could optimise for exploration (novelty) or exploitation (utility), never both. Increasing sampling temperature, tuning nucleus or min-p thresholds, or applying frequency penalties forces a model off its highest-probability tokens, buying surface variation at the cost of reasoning coherence and factual utility.

The strongest published workaround, Regeneration Prompting (sequentially stuffing all prior answers back into the context window and instructing the model to generate something different), partially lifts diversity, but at a punitive cost. Because the prompt grows with every turn, token consumption rises with the square of the number of answers, latency compounds sequentially, and utility still degrades as the context window fills with self-generated noise.

Novamine breaks the trade-off entirely. Operating on an unmodified base model (Claude Opus 5), the Novamine API achieves 9.30/10 distinctness and 6.65 cumulative utility, outperforming Regeneration Prompting even when Regeneration is given a newer, stronger frontier model (Claude Fable 5.1), at a flat token cost that never grows with N.

Outperforming every model and inference-time method on both axes simultaneously.

* Token cost: measured per prompt on our 20-prompt scaling run (Fig. 2). Standard sampling and Novamine need a single 10-answer pass; Regeneration Prompting's 13,043 tokens are its 4-round (N=40) run, because every prior output is fed back into the context window and the footprint compounds with the square of the number of answers.

Why cumulative utility is the decisive metric. NoveltyBench's utility scorer only awards points to answers that are both high-quality and mechanistically distinct from every preceding answer in the set; redundant answers score zero. Regeneration Prompting forces diversity by polluting the context window with prior outputs, paying a steep tax in token overhead and degraded answer quality. The Novamine API achieves near-orthogonal generation (9.30/10) while increasing per-answer quality (6.65), on an older base model, at flat cost.

* Note on the Novamine Full R&D Engine: the public leaderboard entry (9.30 / 6.65) benchmarks our fast, lightweight Elicitation API in isolation. In enterprise R&D deployments, those elicited candidates pass through our Stage 2 Adversarial Hardening Engine, which attacks each mechanism against physical constraints, mathematical derivations and dated prior art before rebuilding survivors. Our independent judge, scoring hardened outputs against the same elicitation layer, places them materially above the API layer; the figures shown are our estimate on the benchmark's scale, not a leaderboard score. Full raw traces are available for technical diligence.

Frontier models know far more than standard inference can extract.

• Intra-model mode collapse (NoveltyBench, CMU, arXiv 2504.05228): standard frontier models sampled 10 times on open-ended problems collapse onto just 2 to 3 distinct ideas, while heuristic decoding tweaks degrade utility.
• Inter-model homogeneity (Artificial Hivemind, Jiang et al., NeurIPS 2025 best paper, arXiv 2510.22954): across 50 independent generations, 79% of open-ended prompts produce near-identical semantic clusters, and different frontier labs converge on the same consensus tropes.
• The alignment bottleneck (ModeBench, Princeton, arXiv 2610.11064): RLHF and post-training reward optimisation narrow output probability mass onto a small subset of "safe" solution strategies. The latent knowledge remains inside the pre-trained weights, but default inference cannot reach it.

Orthogonal to chain-of-thought: Divergent Test-Time Compute

• Depth versus breadth at inference time. Reasoning models (o1, o3, R1, extended thinking) scale test-time compute vertically, spending more tokens deepening a single trajectory down the highest-probability mode. Novamine scales test-time compute horizontally (Divergent Test-Time Compute), steering unmodified reasoning models across mutually exclusive regions of the solution space while holding every trajectory to strict utility floors.
• Why brute-force resampling fails (N=640 empirical proof). A common assumption among ML engineers is that mode collapse can be brute-forced by generating hundreds of samples and filtering them downstream. We tested this directly on Claude Opus 5.5 across 20 NoveltyBench prompts, scaling plain sampling from N=10 to N=640 (631,243 tokens per prompt).
• The receipt (Fig. 2). Even after spending 158× more tokens (N=640) and applying a theoretical oracle selector (using the benchmark's own classifier to cherry-pick the 10 most distinct answers out of 640), plain sampling plateaus at 7.4 distinct ideas, still losing to Novamine's very first 10 answers (8.8 distinct at 3,999 tokens, with zero selection or filtering).

Oracle selection assumes a perfect selector that knows which answers are distinct; none exists outside the scorer. "First 10" is what a user actually receives. Utility was not scored in this run; the leaderboard carries utility.

Measurement one: NoveltyBench, in full

What is measured. 1,100 prompts (100 curated by the authors, 1,000 drawn from real user conversations), each answered ten times, 11,000 generations per entry. Two trained evaluators score every set, neither of them the model under test. Partitioned distinctness is the number of generations among the ten that the benchmark's DeBERTa similarity classifier, calibrated on thousands of human judgements, does not assign to an earlier generation's class. Cumulative utility is scored by a separate reward model and accumulated across the distinct generations only.

How it was listed. We ran the full set through our public endpoint and submitted the outputs. The maintainers re-ran a sample through the same endpoint, reproduced the curated-set numbers themselves, and merged the entry under Inference-Time Algorithms on 10 September 2026. The complete output set and the scoring records are kept as an evidence pack, available on request.

What it does not show. The similarity classifier reads generations in 128-token windows, so very long generations are judged mostly on their openings. The utility evaluator is a reward model, not a domain expert, and rewards general helpfulness rather than correctness in a field. The benchmark does not measure whether an answer is new to the world, only whether it is distinct from the other nine and useful.

Base-model handicap. Our leaderboard submission was executed on Claude Opus 5, whereas the second-place Regeneration Prompting entry was executed on the newer, more capable Claude Fable 5.1. Novamine's +39% novelty and +25% utility advantage was achieved despite running on an older base model.

Measurement two: the Hivemind replication

We replicated Dr Jiang's study with her published instrument on 100 prompts per arm, 50 generations each, then applied Novamine: same models, same prompts, nothing tuned per prompt.

Mode collapse is an elicitation artefact, not a weight limitation. Under Dr Jiang's exact 50-generation protocol on GPT-4o, Novamine reduced mean pairwise semantic similarity from 0.92 to 0.54, collapsing the share of prompts trapped above the 0.80 "Artificial Hivemind" threshold from 96% to 1% (96 of 100 prompts against 1 of 100; Fisher's exact test, p < 10⁻⁵⁰), and transferring zero-shot to Claude Opus 5 without tuning.

Read the replication paper: The Artificial Hivemind, Revisited (PDF) →As sent to Dr Jiang, unchanged.

Her reply, shared on our public submission thread: