honest rebuild · real model calls · no fabricated benchmarks

Test-time compute
that actually runs.

HyperReason is a real Adaptive-Entropy MCTS engine that drives a live language model — with a genuinely interactive demo, a reproducible GSM8K eval, and KV-cache accounting that's labeled projected when it can't be measured.

$ pip install git+github.com/rudra496/hyper-reason
Run the engine ↓
57 tests passing
live GLM-4.6 verified
JS↔Python parity
live in your browser

The engine, running right here.

No API key, no install — the real AE-MCTS loop executes as you watch. It uses a clearly-labeled deterministic sampler so you can see the search mechanics instantly. The pip package runs the same algorithm against a real model. Click any node to inspect its reasoning.

AE-MCTS playground

Ready.
Backend: mock-deterministic-heuristic (is_live=false) — a labeled sampler, not a model. Real reasoning runs in the package. Entropy = sample_diversity_entropy (no logprobs); FlashKV = projected simulator (no real GPU).
reasoning tree · depth →
explore answer unparsable best path
Run a search, then click a node to inspect its reasoning step and statistics.
how it works

Four honest steps.

No hidden heuristics, no hardcoded answers. Every signal comes from real model samples or is explicitly labeled.

① Sample

At each node the model produces K candidate next-steps. Priors and entropy come from those real samples.

② Search

AE-PUCT balances value against sample-diversity entropy — a logprob-free uncertainty proxy. Bounded budget: sims × K calls.

③ Vote

The final answer is majority vote (self-consistency) over terminal trajectories the model actually produced.

④ Account

FlashKV projects CoW paged-cache savings over the real tree — in tokens (model-agnostic) and projected MB.

real benchmark

Published the loss, not a fake win.

Run against the live Z.AI GLM gateway. First N problems of the real GSM8K test split, fixed config, raw per-problem JSONL downloadable — aggregated by a separate script that cannot invent numbers.

greedy T=0, 1 sample
95%
95.0%
self-consistency T=0.7, K=4
90%
90.0%
AE-MCTS sims=6, k=2, d≤3
85%
85.0%

The honest reading: GLM-4.6 is already near the ceiling on easy GSM8K, so a modest-budget search underperforms greedy here. Test-time search pays off on harder problems / weaker base models / bigger budgets — not on a strong model at its ease. Not a claim of SOTA. Re-run it: python eval/gsm8k_mini.py --n 100 --sims 16 --k 4 · Raw: eval/runs/*.jsonl

the honesty contract

What's real, and what's labeled.

v1.x shipped a wrapper that never called a model and a benchmark that was invented. v2 is bound by these rules.

rule 1

Real or labeled

Every number is produced by a runnable command (config given) or explicitly tagged projected / simulated with the assumption stated.

rule 2

No fake fallbacks

A missing model or daemon makes the backend raise — never return invented text or token counts. (v1.x silently faked Ollama.)

rule 3

Entropy, stated

A sample-based uncertainty proxy labeled sample-diversity entropy (no logprobs). The gateway returns no per-token distribution.

rule 4

Verifiable

Raw per-problem JSONL + a separate aggregator in eval/; a JS↔Python parity test pins the browser demo to the package.