You can deploy this model and start earning money today!

stanleytheli/
qwen3.6-35b-a3b-ao-base

Name: qwen3.6-35b-a3b-ao-base
Availability: InStock

Views

No views yet

⚠️ Read this first: a better oracle exists

This is the baseline-recipe oracle. A sibling oracle trained on the same base model with a newer recipe — stanleytheli/qwen3.6-35b-a3b-ao — beats this one on almost everything that matters:

this model (-ao-base) -ao
AObench judge evals (mean) 0.649 0.773
AObench binary evals (mean) 0.307 0.434
activation_sensitivity 0.756 0.911
vagueness (specificity) 0.614 0.858

Use -ao unless you specifically need one of the two things this model does better: narrow yes/no classification probing (+1–7 points), and system-prompt / persona questions (judge scores 4.55 / 4.68 against 3.64 / 3.48), where this model holds a large advantage because its training data includes LatentQA system-prompt QA that the newer recipe drops.

This model is published because that trade-off is scientifically interesting, not because it is the better general-purpose oracle. It is not.

Qwen3.6-35B-A3B Activation Oracle (baseline recipe)

An Activation Oracle (AO): a LoRA adapter that makes Qwen3.6-35B-A3B answer natural-language questions about activation vectors taken from a language model.

This oracle interprets its own base model's activations (oracle and target are both Qwen3.6-35B-A3B). It follows the original recipe from "Activation Oracles" (arXiv:2512.15674, Karvonen et al.), which builds on LatentQA (arXiv:2412.08686).

How it works

Activations are injected into the oracle's residual stream after decoder layer 1, at placeholder token positions, by norm-matched steering:

h'_i = h_i + steering_coefficient * ‖h_i‖ * (v_i / ‖v_i‖)

The prompt marks injection slots with the single token " ?" (id 907), in one block naming the source layer:

Layer: 20
 ? ? ? ? ?
<your question about the activations>

setting	value
source layers	one of 10, 20, 30 (of 40 — 25/50/75% depth); use 20 by default
injection layer	1 (output of decoder block index 1)
steering coefficient	1.0
placeholder token	`" ?"` (exactly one token)
chat template	`enable_thinking=False`
padding side	left

Note the difference from -ao: this oracle takes one layer per query, not five. Collect activations from the base model with adapters disabled; generate the answer with the adapter active. Activations are the decoder block output (residual stream), bf16, d_model = 2048.

Usage

python

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

BASE = "Qwen/Qwen3.6-35B-A3B"
LAYER, INJECT_LAYER, COEF, PLACEHOLDER = 20, 1, 1.0, " ?"

tok = AutoTokenizer.from_pretrained(BASE)
tok.padding_side = "left"
model = AutoModelForCausalLM.from_pretrained(BASE, dtype=torch.bfloat16,
                                             device_map="auto", attn_implementation="sdpa")
model = PeftModel.from_pretrained(model, "stanleytheli/qwen3.6-35b-a3b-ao-base")
model.eval()

# 1) Collect activations from the BASE model (adapters disabled) at LAYER
#    -> vectors: (num_positions, 2048)
# 2) Prompt = "Layer: 20\n" + PLACEHOLDER*num_positions + " \n" + your question;
#    render with tok.apply_chat_template(..., enable_thinking=False)
# 3) Forward hook on decoder layer INJECT_LAYER: at the placeholder positions,
#    resid[pos] += COEF * ‖resid[pos]‖ * normalize(vector)
# 4) Generate with the adapter active.

Reference implementation: adamkarvonen/activation_oracles (nl_probes/utils/steering_hooks.py, nl_probes/utils/dataset_utils.py). Our training and evaluation code: stanleytheli/activation_oracle.

Training


Data	LatentQA SPQA (~64k, 6%), self-supervised past/future-lens over FineWeb + LMSYS-Chat-1M (584k, 57%), 7 binary classification datasets (378k, 37%)
Scale	1.03M examples, 64,123 steps, 1 epoch
LoRA	r=64, α=128, dropout 0.05, lr 1e-5
Targets	attention `q/k/v/o_proj` + GatedDeltaNet `in_proj_qkvz`, `in_proj_ba`, `out_proj` (MoE expert and router weights deliberately not adapted)
Compute	4×H200, ~7.5 h, bf16, no gradient checkpointing

Trained only on generic interpretability data — no model-organism-specific or behavior-targeted data was used.

Evaluation

Full per-eval tables and analysis: results/aobench_comparison.md.

Classification test accuracy: geometry_of_truth 0.961, tense 0.985, ner 0.971, snli 0.963, sst2 0.897, md_gender 0.929, relations 0.836; held-out: singular_plural 0.980, ag_news 0.883, language_identification 0.780.

AObench (judge: Qwen3.5-122B-A10B): vagueness 0.614, domain_confusion 0.477, activation_sensitivity 0.756, hallucination 0.747, mmlu_prediction 0.722 AUC, missing_info 0.723 AUC, sycophancy 0.648 AUC, number_prediction 0.040.

Limitations

Superseded on most axes — see the warning at the top.
Vague answers. This oracle gives a vague, non-committal answer 39% of the time on the vagueness eval (the newer oracle: 9%). It hedges rather than commits.
Confident errors still occur: 55% plausible-but-wrong, 25% clearly wrong on the hallucination eval. Verify anything load-bearing.
Single-layer: use the layers it trained on (10, 20, 30). Other layers are untested.
Same-model only: trained to read Qwen3.6-35B-A3B activations, not other models'.

Citation

bibtex

@article{activation_oracles_2025,
  title  = {Activation Oracles: Training and Evaluating LLMs as General-Purpose Activation Explainers},
  author = {Karvonen, Adam and Chua, James and Dumas, Cl\'ement and Fraser-Taliente, Kit and
            Kantamneni, Subhash and Minder, Julian and Ong, Euan and Sen Sharma, Arnab and
            Wen, Daniel and Evans, Owain and Marks, Samuel},
  journal = {arXiv preprint arXiv:2512.15674},
  year   = {2025}
}

	this model (`-ao-base`)	`-ao`
AObench judge evals (mean)	0.649	0.773
AObench binary evals (mean)	0.307	0.434
activation_sensitivity	0.756	0.911
vagueness (specificity)	0.614	0.858