The consensus bottleneck, and how we break it

01
The problem

Leading AI converges on the consensus answer.

1

Every generation of AI has collapsed toward the consensus

Every generation of generative AI has gravitated toward its most probable output. “Mode collapse” was named in 2016 to describe how image generators homogenised, years before ChatGPT.

There is still no complete fix at the training stage: training needs a target to score against, and “different but still right” is far harder to score than “right”.

Goodfellow, NIPS 2016 GAN tutorial; Salimans et al., NeurIPS 2016
2

Ask four leading AI models an open-ended creative question

YOUGive me five names for my new band.
AI 1
Velvet EchoNeon TideMidnight ArcadeStatic BloomThe Hollow Pines
AI 2
Neon VelvetEcho ChamberMidnight StaticGolden HourPaper Moons
AI 3
The Velvet HoursMidnight EchoesNeon WolvesLunar TideSilver Static
AI 4
Echo ValleyVelvet StaticNeon MidnightWild CoastCopper Skies

Four independent models reach for the same few words.

Typical answers, shown for illustration
3a

Now ask the same four a scientific R&D question

SCIENCEHow could we stop antibiotic resistance spreading?

All four land in the same narrow textbook cluster.

Individually plausible, collectively homogeneous. Same colour = same answer. Typical answers, for illustration · Artificial Hivemind: Jiang et al., NeurIPS 2025
3b

Or a commercial strategy question

STRATEGYHow does a UK building society win under-30 savers without raising rates?

The same strategic orthodoxy, whichever company asks.

Same colour = same answer. Typical answers, for illustration · Artificial Hivemind: Jiang et al., NeurIPS 2025
4

The commercial cost: no differentiation

When your R&D and strategy teams use the same models as your competitors, everyone converges on the same crowded ideas. No edge, and no breakthrough.

Different models give strikingly similar answers: Jiang et al., NeurIPS 2025
5

Your work is converging with your competitors’

Memos, proposals and pitches drafted with standard AI read well on their own, yet converge on the same theses across the market. Inside one firm the collapse is invisible; to clients and investors it is unmistakable.

AI raised individual quality while reducing collective diversity: Doshi & Hauser, Science Advances 2024
02
Why it happens

Training and tuning narrow the range.

1 · Training

AI learns from AI-written text

New models learn from text that earlier models wrote. The rarer answers drop out.

2 · Tuning

Raters reward the expected

People, and now other AIs, rate the answers. The familiar one scores best, so the model learns to give it.

3 · Locked in

Fixed before you use it

The narrowing is built into the finished model. Rewording your prompt does not undo it.

Model collapse: Shumailov et al., Nature 2024 · typicality bias: Zhang et al., 2025 · NoveltyBench: Zhang et al., 2025

03
What Novamine does

We unlock the full solution space.

Distinct answers across the whole solution space, each heavily stress-tested so you don’t have to. Here is what comes back.

1a

The same scientific R&D question, asked of Novamine

SCIENCEHow could we stop antibiotic resistance spreading?
Six distinct mechanisms, none from the textbook cluster. Different colour = different answer. Real Novacore output from our logged runs, each condensed to one line.
1b

And the same commercial strategy question

STRATEGYHow does a UK building society win under-30 savers without raising rates?
Six distinct moves, none on the standard list. Different colour = different answer. Real Novacore output from our logged runs, each condensed to one line.
Route 1 · The full Novamine engine (Collaborative Deep-Dive)
The intractable R&D or strategy problemThe full agentic adversarial network, run multiple times over 24 to 72 hours, as directed by both your experts and ours working collaboratively.
What you get
1

A summary dossier of up to 10 IP-ready theses

One page per thesis, grouped by the kind of bet each makes, from defensible flagships to bolder leaps.

  • A composite score out of 10, with separate marks for novelty, defensibility, the decisiveness of the first test and adherence to your brief
  • The orthodox approach, set against the Novamine leap
  • Build and first test: the experiment or pilot that settles it
  • Risks and limitations, stated plainly
2

The full paper behind every thesis

Up to ten papers, one per thesis, each written to stand on its own in front of your technical leads.

  • Your brief, constraints, forbidden routes and stage gates
  • The insight, with a non-technical explanation
  • The leap, against the orthodox approach
  • Numeric derivation
  • Parts list and build instructions
  • Cost and resource estimates
  • Compliance and regulatory considerations
  • Prior art status
  • Intellectual honesty: our own caveats
  • Emergent questions and new frontiers

Each thesis is checked against prior art and, as far as our searches can establish, is without precedent. Composite scores of around 9 out of 10 are typical. The IP is yours.

Route 2 · The Novacore API
Divergence inside your own stackA drop-in step (3 to 60 minutes, Rigour 0 to 3) at the point where an existing workflow needs non-consensus breadth. An API (Application Programming Interface) your team plugs directly into your existing systems.
What you get
  • Every response ranked and scored for novelty and utility
  • Checked against what already exists
  • You choose how novel and how rigorous
  • Your data is never used for training
  • Ranked #1 on CMU NoveltyBench (9.30/10 distinctness · 6.65 utility)
In your organisation

Enterprise AI assistant: a ranked candidate set

Non-overlapping answers returned inside your own assistant, each scored for novelty and utility. The user picks a direction and carries on, context intact.

Your R&D function

Discovery workflow: screened, distinct hypotheses

Mechanistically distinct, screened against prior art and scored. Your in-silico triage and wet-lab validation proceed unchanged.

Your planning process

Strategy workflow: stress-tested options

Distinct strategies, each challenged and scored against the brief. Your team runs the financials on the strongest and takes one to the board.

High-throughput uses of the same API
Agent stacks

Answer pools for multi-agent and forecasting stacks

Independent reasoning paths for mixture-of-agents, debate and forecasting ensembles. Diversity is what makes a crowd accurate: crowd error = average error − diversity Diversity Prediction Theorem: Page, The Difference, 2007

Model training

RLAIF and synthetic training data

Scored, distinct responses for RLAIF tuning and evaluation sets, so your next model learns from the full range, not the collapsed middle.

Your own data

Retrieval-augmented generation (RAG) over your own data

Your retrieval layer pulls the context from your knowledge base, data lake or warehouse; Novacore replaces the single consensus answer with distinct, scored readings of the same evidence, so mode collapse does not follow you into your own data.