AUBIN-12B-Control — AUBIN by Norovox

Real-time control adapter for AubinController: observation (text) + command → one move with a calibrated confidence, one forward pass per move. Part of the AUBIN family (see emrevrg/AUBIN-12B).

With the same safety shield, AUBIN-12B-Control reaches 92% success with 0 lava deaths, vs 89% for the strongest rule baseline (greedy + shield) on the same 100 unseen episodes. An untrained Gemma-4-E4B base without control training reaches 11% with the shield.

Benchmark: command-following grid game (100 episodes, seed 0, never seen in training)

7x7 grid as text, walls, 5 deadly lava cells, 4 objects that block movement, command like "go to the red key". Step accuracy = the move is on a shortest safe path (BFS ground truth). Training used 3,000 procedurally generated episodes (seeds 1000+, 15,161 states, soft labels over all shortest moves); the test episodes are disjoint. The safety shield only removes moves into visible lava/walls/objects (AubinController.act(..., allowed=safe_moves)); the same shield is applied to the rule baselines so its effect is separated from the model's.

controller	success	lava deaths ↓	step accuracy	median ms / move (T4, 4-bit)
random	6%	40	35.3%	–
random + safety shield	13%	0	53.4%	–
greedy (ignores obstacles)	58%	26	56.8%	–
greedy + safety shield	89%	0	87.6%	–
AUBIN-12B-Control (one pass per move)	74%	24	86.4%	854
AUBIN-12B-Control + safety shield (never steps on visible lava/walls)	92%	0	91.8%	856
Gemma-4-E4B base, no control training (one pass per move)	8%	26	39.2%	329
Gemma-4-E4B base, no control training + safety shield	11%	0	55.4%	330

Use

pip install "git+https://huggingface.co/emrevrg/AUBIN-12B#subdirectory=code"
from aubin import Aubin, AubinController
ctl = AubinController(Aubin("emrevrg/AUBIN-12B-Control"),
                      actions={"up": "move one cell up (row - 1)", "down": "move one cell down (row + 1)",
                               "left": "move one cell left (column - 1)", "right": "move one cell right (column + 1)"},
                      instructions="Choose the next move that follows the command along a shortest safe path "
                                   "(never step on lava, objects block movement).")
ctl.reset()                                    # new episode
r = ctl.act(observation, command="go to the red key", allowed={"up", "left"})   # allowed = moves that are safe right now
r["action"], r["confidence"], r["latency_ms"]

Honest notes

• Without the shield the adapter alone is not better than a greedy rule on this game (see table); its value shows when it chooses among safe moves. The shield is a standard action mask, not a planner: it never looks ahead.
• Latency is measured on a free Colab T4 in 4-bit, batch 1. Benchmark code: control_bench.py, training: control_train.py (in the AUBIN-12B repo code/).
• Base model: google/gemma-4-12B-it (Apache-2.0). This repo holds only the LoRA adapter (fp16) and its measurement file.

AUBIN-Learn — instant self-learning (experimental, Norovox core)

AUBIN can learn from feedback without retraining. Verified cases go into an external decision memory (a write takes well under a millisecond with the hashing embedder), and a self-calibrator (Hedge / multiplicative weights) shifts trust per source between the model, the memory and their fusion — so where memory is not useful yet, AUBIN keeps trusting itself. Measured on Kev's public suites; full tables, protocol and code: reports/AUBIN_LEARN.md, code/aubin/learn.py.

• Feedback stream, kev_test: all 6 AUBIN variants improve, +1.0 to +1.3 points (best 84.3 → 85.6). Separate protocol — the label is revealed after each answer — so it is not comparable to static scores (Kev-9B 87.4 static).
• Never-seen sources (transfer, memory starts empty): −0.4 to +0.1 points — it does not hurt; a few hundred feedbacks per source are not enough to help yet.
• Fast skills, static locked test: per-source classifiers learned from memory in seconds (switched on only where dev proves them) lift kev_test for all 4 measured AUBIN variants, +0.3 to +0.6 points (ensemble 84.3 → 84.8; banking77 59.5 → 65.5).
• Raw memory (kNN) on the static test: no reliable gain (−1.0 to +0.7) — AUBIN already learned these sources.

Results, 3 October 2026 (full report: reports/AUBIN_RESULTS_2026-10-03.md)

benchmark	AUBIN	reference
Typed decisions (2,000 decisions)	77.55 (AUBIN-Learn, weights fixed before test)	Laya 76.65 · meraGPT 76.8 · Jev 72.7
Kev suites, Kev's training sources (kev_test)	85.7 (AUBIN ensemble, selected on cal split)	Kev-0.8B 83.8 · Kev-4B 86.5 · Kev-9B 87.4
Kev suites, transfer test	86.5 ensemble · 89.0 AUBIN-31B	–
Mind2Web cross-domain step SR (200 steps)	48.5 AUBIN-31B, no web training	MindAct-XL 39.6 · GPT-4 26.4
Grid control, 100 unseen episodes	92% success, 0 lava deaths (12B-Control + shield)	greedy rule + same shield 89%

On Kev's own training sources AUBIN is still 1.7 points behind Kev-9B; this is stated, not hidden.