release: v0.1.1 — enable_thinking=False default + corrected bench gold + CHANGELOG

CHANGELOG.md

@@ -0,0 +1,18 @@
+# Changelog
+
+## v0.1.1 — 2026-05-10
+
+### Added
+- `RSAConfig.enable_thinking` field (default `False`). Toggling `<think>` mode in the chat template substantially affects output quality on math problems.
+- Bench `scripts/bench_hmmt.py` now uses corrected gold answers for the placeholder HMMT-1 (66, was 100) and HMMT-5 (1, was 76).
+
+### Changed
+- Default `enable_thinking` flipped to `False`. Empirical testing shows `<think>` mode causes the model to narrate the aggregation prompt (`"We have a user message: ..."`) instead of solving. Direct mode produces math reasoning immediately.
+- `_render_chat(messages, *, enable_thinking)` signature now takes an explicit kwarg (was hardcoded to `True`).
+
+### Bench results
+- 5/5 vanilla + 5/5 RSA on corrected HMMT subset. lift_pp +0.00pp (ceiling effect — vanilla already at 100%).
+
+## v0.1.0 — 2026-05-10
+
+Initial public release. T=2 N=4 RSA orchestrator with audit JSONL + CLI + HMMT bench harness.

README.md

@@ -2,7 +2,7 @@
 
 First MLX implementation of Zyphra's **Markovian RSA** test-time compute methodology, targeting **ZAYA1-8B** on Apple Silicon. Boosts reasoning accuracy by sampling N parallel reasoning traces, extracting their tails, and feeding aggregation prompts back to the model.
 
-> **Status :** v0.1.0. Aggregation prompt is `zaya_v1` (reverse-engineered ; paper does not publish the co-trained format). HMMT'25 5-problem smoke shows ≥ 0 pp lift on M2 Pro.
+> **Status :** v0.1.1. `enable_thinking=False` default ; aggregation `zaya_v1` template (reverse-engineered ; paper does not publish co-trained format). Both vanilla and RSA score 100% on the 5-problem corrected HMMT subset (ceiling effect — needs harder set for real lift measurement).
 
 ## Install
 
@@ -43,6 +43,17 @@ markovian-rsa-mlx solve "Compute the integral of x^2 from 0 to 5" \
 | `paper-16k` | 2 | 4 | 16 K | ~ 16-24 GB | paper "deployment" profile |
 | `paper-headline-40k` | 2 | 16 | 40 K | 32+ GB | paper headline (HMMT'25 89.6) |
 
+## Bench results (HMMT'25 5-problem subset)
+
+With the corrected placeholder dataset and `enable_thinking=False` default :
+
+| Backend | Score | Wall time | Per-problem avg |
+|---|---:|---:|---:|
+| Vanilla (T=1 N=1) | 5/5 = 100% | 1085 s | 217 s |
+| RSA T=2 N=2 (default-16gb) | 5/5 = 100% | 3974 s | 795 s |
+
+`lift_pp = +0.00pp` on this subset due to ceiling effect (vanilla already hits 100%). Larger HMMT'25 / AIME'26 datasets needed to measure the real lift. The system is mechanically correct (RSA outputs reference "Approach 1, Approach 2" from aggregation prompts) ; just needs harder problems to differentiate.
+
 ## Audit JSONL
 
 Every event of the run is one line. Schema in

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "markovian-rsa-mlx"
-version = "0.1.0"
+version = "0.1.1"
 description = "Markovian RSA test-time compute methodology on MLX for ZAYA1-8B and future co-trained models"
 readme = "README.md"
 requires-python = ">=3.12,<3.14"

src/markovian_rsa_mlx/__init__.py

@@ -1,5 +1,5 @@
 """Markovian RSA test-time compute methodology on MLX."""
-__version__ = "0.1.0"
+__version__ = "0.1.1"
 
 from markovian_rsa_mlx.config import RSAConfig
 from markovian_rsa_mlx.loader import load_zaya_model

tests/test_cli.py

@@ -7,7 +7,7 @@ runner = CliRunner()
 def test_version_command_prints_version():
     result = runner.invoke(app, ["version"])
     assert result.exit_code == 0
-    assert "0.1.0" in result.stdout
+    assert "0.1.1" in result.stdout
 
 
 def test_solve_help_shows_required_flags():

uv.lock

@@ -421,7 +421,7 @@ wheels = [
 
 [[package]]
 name = "markovian-rsa-mlx"
-version = "0.1.0"
+version = "0.1.1"
 source = { editable = "." }
 dependencies = [
     { name = "huggingface-hub" },