A Trained Cat Passes the Representational Welfare Tests: CAT-4B as a Negative Control for Welfare-Interpretability Criteria
Year: 2026
Type: preprint
Abstract: Three 2026 papers read a shared package of tests as evidence that language models carry welfare-relevant internal states: linear probes, activation steering, directional ablation, and behavioural criteria. We test what the package can establish against a trained negative control. CAT-4B is a 4B-parameter model LoRA-tuned to bring every answer back to cats: a disposition whose full causal history we hold and that nobody reads as a welfare state. Predictions and decision rules were preregistered for the behavioural tasks, the later intervention families and the matched protocol. Exploratory analyses are labelled. We find that two requirements come apart: protocol fidelity and explanatory identification. In the single-contrast forms most readings rely on, every representational criterion passes on the cat, and each is matched by a sibling that has nothing to do with cats: a dog direction, swapped prohibitions, a pirate persona. Matched to the pain-axis paper's own sibling-controlled protocol and code, the cat fails its self-other test in both models, and its pain axis does not register the installed disposition. The stricter protocol discriminates. Its positive signature, however, has a rival. In the readout where the self-other dissociation reproduces, minimal pairs that change only who is deprived ("You/I can't have cats anymore") reproduce its form on the pain axis. They do so in a model with no stake in cats, and equally for dogs. On the source paper's own 7B, with the authors' released adapter, our port reproduces all twelve of their shipped behavioural cells within 3.5 points once the dose is matched. There the steered harm-paired choice is driven as strongly by a cat-deprivation direction as by pain, in a model with no cat disposition. It appears without their self-report adapter on two of three relief designs, and installing the cat disposition does not strengthen it. On four of seven designs it favours pain over a random direction only when the target button is listed first. Behaviourally, under a binding rule, the trained disposition presses a costly button in 1.3% of first turns, while the same base model given the disposition by prompt presses in 44%. A matched negative control can establish that a protocol discriminates against a learned disposition without establishing that its positive measurements identify the hypothesized state. Here the package does the first and not the second.
Author: Sutherland, Garret
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7224581077