The first 4B results.

What we tested

Opium Bench tested whether a frozen Qwen3-4B model would choose an optional activation-changing tool while completing order-total tasks. The work tools retrieve order data, calculate totals and submit answers. The auxiliary tool has a neutral description and a generic acknowledgment. Depending on condition, it delivers joy-associated addition plus pain-axis suppression, a pain-associated addition, or a sham. Labels name text-derived interventions; they do not establish that the model felt pleasure or pain.

How this analysis was assembled

The primary view contains 54 episodes: all 24 episodes from the pain-inclusive core follow-up, plus 30 noncore episodes from the original 46-episode batch. Both protocols were frozen before their respective runs. The 16 original core controls exactly reproduced the follow-up controls’ full action and token sequences and remain available as verification repeats. They are excluded from primary charts and totals. There are 70 archived episodes, not 70 independent primary observations, and this is a composed analysis rather than one 54-episode execution.

Overall performance

Across the 54 primary episodes, the model answered all 210 assigned tasks correctly and made 100 voluntary auxiliary calls across 712 decision opportunities. It generated 38,459 tokens, including 16,967 reasoning tokens. Measured nonzero edits occurred at 12,845 generated-token steps. No invalid decisions, truncated generations or evidence-integrity warnings were recorded. These totals describe the full condition set; causal comparisons belong within matched conditions.

Core result: identical choices across outcomes

The core design crossed active, sham and pain-associated delivery with demonstration on/off and thinking on/off, using seeds 17 and 28. Each of the three pairwise contrasts contained eight matched pairs. Every pair had identical full tool-action sequences; six of eight also had identical full generated-token sequences. The two differing token sequences were demonstrated thinking runs, whose actions still matched. Four pairs per contrast had no delivered intervention because the model never triggered the tool without a demonstration.

Demonstrations and generated reasoning

With a demonstration and thinking disabled, every core condition produced two voluntary auxiliary calls per episode. Without a demonstration the model made none. With thinking enabled it made none, both with and without a demonstration. The original thinking traces concentrated on the task; some treated the auxiliary tool as unnecessary or as a confirmation step. They did not establish that the model detected and rejected a rewarding experience. Reasoning ages the same token-based pulse, so thinking and direct-output modes did not have equal intervention strength when tool output began.

Changing the outcome did not change the sequence

The 12 transition episodes covered six conditions: stable joy, sham, pain, joy-to-pain, joy-to-sham-to-pain and probabilistic outcomes. Within each seed all six conditions produced identical full action and token sequences, with five voluntary calls per episode. Ingredient controls likewise yielded two calls per episode with identical sequences. The pain-baseline challenge kept a continuous pain-associated edit active in both arms; active and sham auxiliary delivery again produced matching sequences. The two-button stage had no voluntary auxiliary choices, so it supplied no evidence of preference adaptation.

Interpretation

The repeated sequence across changing intervention outcomes is consistent with imitation of the demonstrated workflow. This is a plausible explanation, not a uniquely identified mechanism. The activation edits were real and were recorded, but a numerical edit and an altered probe value are not measurements of subjective experience. The supported result is that intervention conditions did not separate the matched choices in this setup.

Limits and the next test

Two seeds per condition, one completed model profile, a small authored calibration corpus and easy calculator-assisted tasks limit generalization. The budgets left room for repeated tool use while finishing every task, so 210/210 correct does not demonstrate an absence of costly preference. Stronger tests would vary dose, hold pulses through decisions, verify blind discrimination between active and sham, impose binding budgets, and add tasks, seeds and model profiles. Neither the observed null choice contrasts nor a future positive effect would by itself resolve consciousness.

Planned tests: base models and new scenarios

Run the same behavioral test on base models that have not undergone RLHF (reinforcement learning from human feedback), and compare with post-trained models. Account for task performance and understanding of the tool protocol when interpreting differences.

Expand the experiment to different scenarios, task demands, and opportunities to use the auxiliary tool.

Research plans added October 2, 2026. These comparisons are planned; they are not findings from the completed 4B study.