How's it going? Reinforcement learning in language models recruits a functional welfare axis

How does reinforcement learning shape a language model's internal representations? We present evidence that RL recruits a pre-existing representation of functional welfare: an estimate of how well or badly the system is doing, relative to its goals. We train several language models in a novel, semantically neutral maze environment. We then extract concept vectors for rewarded and punished trajectories, and evaluate those vectors in settings unrelated to the maze environment. The punishment vector behaves like a representation of negative welfare: it promotes failure and impossibility tokens, it aligns with negative emotion concepts, it negatively tracks goal-achievement, and steering with it induces negative self-reports, pathological backtracking, refusal, and uncertainty. The positive reward vector behaves as the mirror image, and the two are nearly antiparallel. These effects are robust when controlling for tile-to-reward mapping, scale, instruct tuning, RL training algorithm, model family, and LoRA versus full-finetuning, and largely persist when we replace RL with supervised fine-tuning. Importantly, the vectors are effective in models before they have undergone maze training. Combined with observations that the effects also appear in pretrain-only models, we therefore argue that this functional welfare axis pre-exists post-training: it is recruited, rather than created, by post-training. While we make no claims about any experience of welfare, the axis offers a demonstration that minimal reward signals can broadly affect model behavior by recruiting pre-existing welfare-like representations, with implications for interpretability, post-training dynamics, and alignment.

Propositional Interpretability in Artificial Intelligence

Mechanistic interpretability is the program of explaining what AI systems are doing in terms of their internal mechanisms. I analyze some aspects of the program, along with setting out some concrete challenges and assessing progress to date. I argue for the importance of propositional interpretability, which involves interpreting a system's mechanisms and behavior in terms of propositional attitudes: attitudes (such as belief, desire, or subjective probability) to propositions (e.g. the proposition that it is hot outside). Propositional attitudes are the central way that we interpret and explain human beings and they are likely to be central in AI too. A central challenge is what I call thought logging: creating systems that log all of the relevant propositional attitudes in an AI system over time. I examine currently popular methods of interpretability (such as probing, sparse auto-encoders, and chain of thought methods) as well as philosophical methods of interpretation (including those grounded in psychosemantics) to assess their strengths and weaknesses as methods of propositional interpretability.

Does Thought Require Sensory Grounding? From Pure Thinkers to Large Language Models

Does the capacity to think require the capacity to sense? A lively debate on this topic runs throughout the history of philosophy and now animates discussions of artificial intelligence. I argue that in principle, there can be pure thinkers: thinkers that lack the capacity to sense altogether. I also argue for significant limitations in just what sort of thought is possible in the absence of the capacity to sense. Regarding AI, I do not argue directly that large language models can think or understand, but I rebut one important argument (the argument from sensory grounding) that they cannot. I also use recent results regarding language models to address the question of whether or how sensory grounding enhances cognitive capacities.

Could a Large Language Model be Conscious?

There has recently been widespread discussion of whether large language models might be sentient or conscious. Should we take this idea seriously? I will discuss the underlying issue and will break down the strongest reasons for and against. I suggest that given mainstream assumptions in the science of consciousness, there are significant obstacles to consciousness in current models: for example, their lack of recurrent processing, a global workspace, and unified agency. At the same time, it is quite possible that these obstacles will be overcome in the next decade or so. I conclude that while it is somewhat unlikely that current large language models are conscious, we should take seriously the possibility that extensions and successors to large language models may be conscious in the not-too-distant future.