iiinova — where your digital self keeps evolving

Raphaël Millière @raphaelmilliere.com · 14/04/2026

Looking forward to speaking at this ICML workshop on ML & Philosophy! Check out the full lineup and CFP below (deadline: May 11th). Despite the title, the CFP is open to work in many areas of the philosophy of AI, not just AI ethics. sites.google.com/view/philmli...

1132

Raphaël Millière @raphaelmilliere.com · 20/02/2026

Thanks! We have a fully updated review paper on this forthcoming in Philosophy Compass (should be preprinted soon), and a book in preparation that gets into more details and could be used as a textbook for that kind of course.

160

Raphaël Millière @raphaelmilliere.com · 02/02/2026

Very glad to see this out! Great paper

Tomer Ullman @tomerullman.bsky.social · 27/01/2026

Reposted by Raphaël Millière

now accepted at ICLR! 🐺🥳🐺 arxiv.org/abs/2506.20666

0409

Raphaël Millière @raphaelmilliere.com · 20/01/2026

The main takeaway for me is that structural information in language is far more constraining than intuition suggests. That's very interesting (and I agree that parrot metaphors are misleading) but it seems like a claim about language more than intelligence. 3/3

1210

Raphaël Millière @raphaelmilliere.com · 20/01/2026

The LLM has to do something like schema-conditioned infilling: produce a high-probability member of the equivalence class consistent with those constraints. So I'm not sure how unexpected the results are? That's roughly what I'd expect from matching to structurally similar training passages. 2/3

190

Raphaël Millière @raphaelmilliere.com · 20/01/2026

Very cool idea! Some quick thoughts. It looks like the corruption preserves a lot of information (function words, morphology, word order, punctuation, numbers, register) which would strongly constrains the posterior over plausible discourse frames as it were. 1/3

1171

Reposted by Raphaël Millière

Department of Statistics @oxfordstatistics.bsky.social · 26/08/2025

With @jesusoxford.bsky.social we are looking for a Professor of Statistics. Become part of a historic institution and a community focused on academic excellence, innovative thinking, and significant practical application. About the role: tinyurl.com/b8uy6mr5 Deadline: 15 September

143

Raphaël Millière @raphaelmilliere.com · 21/08/2025

I'm happy to share that I'll be joining Oxford this fall as an associate professor, as well as a fellow of @jesusoxford.bsky.social and affiliate with the Institute for Ethics in AI. I'll also begin my AI2050 Fellowship from @schmidtsciences.bsky.social there. Looking forward to getting started!

3480

Raphaël Millière @raphaelmilliere.com · 14/08/2025

Thanks Ali! We'll (hopefully soon) publish a Philosophy Compass review and for a longer read a Cambridge Elements book that are the spiritual successors to these preprints and up-to-date w/ both technical and philosophical recent developments

This opens intesting avenues for future work. By using causal intervention methods with open-weights models, we can start to reverse-engineer these emergent analogical abilities and compare the discovered mecanisms to computational models of analogical reasoning. 8/9

1190

But models also showed different sensitivities than humans. For example, top LLMs were more affected by permuting the order of examples and were more distracted by irrelevant semantic information, hinting at different underlying mechanisms. 7/9

1221

We found that the best-performing LLMs match human performance across many of our challenging new tasks. This provides evidence that sophisticated analogical reasoning can emerge from domain-general learning, where existing computational models fall short. 6/9

1172

In our second study, we highlighted the role of semantic content. Here, the task required identifying specific properties of concepts (e.g., "Is it a mammal?", "How many legs does it have?") and mapping them to features of the symbol strings. 5/9

1150

In our first study, we tested whether LLMs could map semantic relationships between concepts to symbolic patterns. We included controls such as permuting the order of examples or adding semantic distractors to test for robustness and content effects (see full list below). 4/9

1160

We tested humans & LLMs on analogical reasoning tasks that involve flexible re-representation. We strived to apply best practices from cognitive science – designing novel tasks to avoid data contamination, including careful controls, and doing proper statistical analysis. 3/9

1140

We focus on an important feature of analogical reasoning often called "re-representation" – the ability to dynamically select which features of analogs matter to make sense of the analogy (e.g. if one analog is "horse", which properties of horses does the analogy rely on?). 2/9

1190

Raphaël Millière @raphaelmilliere.com · 18/07/2025

Can LLMs reason by analogy like humans? We investigate this question in a new paper published in the Journal of Memory and Language (link below). This was a long-running but very rewarding project. Here are a few thoughts on our methodology and main findings. 1/9

514038

I'm glad this can be useful! And I totally agree regarding QKV vectors – focusing on information movement across token positions is way more intuitive. I had to simplify things quite a bit, but hopefully the video animation is helpful too.

010

Raphaël Millière @raphaelmilliere.com · 18/07/2025

See also @melaniemitchell.bsky.social's excellent entry on Large Language Models: oecs.mit.edu/pub/zp5n8ivs/

oecs.mit.edu

Large Language Models

0112

Raphaël Millière @raphaelmilliere.com · 18/07/2025

I wrote an entry on Transformers for the Open Encyclopedia of Cognitive Science (‪@oecs-bot.bsky.social‬). I had to work with a tight word limit, but I hope it's useful as a short introduction for students and researchers who don't work on machine learning: oecs.mit.edu/pub/ppxhxe2b

oecs.mit.edu

Transformers

15412

Raphaël Millière @raphaelmilliere.com · 14/07/2025

Happy to share this updated Stanford Encyclopedia of Philosophy entry on 'Associationist Theories of Thought' with @ericman.bsky.social. Among other things, we included a new major section on reinforcement learning. Many thanks to Eric for bringing me on board! plato.stanford.edu/entries/asso...

plato.stanford.edu

Associationist Theories of Thought (Stanford Encyclopedia of Philosophy)

1398

Reposted by Raphaël Millière

Sam Gershman @gershbrain.bsky.social · 09/07/2025

The sycophantic tone of ChatGPT always sounded familiar, and then I recognized where I'd heard it before: author response letters to reviewer comments. "You're exactly right, that's a great point!" "Thank you so much for this insight!" Also how it always agrees even when it contradicts itself.

518722

Normative conflicts and shallow AI alignment - Philosophical Studies

The paper is available in open access. It includes a lot more, including a discussion of how social engineering attacks on humans relate to the exploitation of normative conflicts in LLMs, and some examples of "thought injection attacks" on RLMs. 13/13 link.springer.com/article/10.1...

link.springer.com

The progress of AI systems such as large language models (LLMs) raises increasingly pressing concerns about their safe deployment. This paper examines the value alignment problem for LLMs, arguing tha...

060

In sum: the vulnerability of LLMs to adversarial attacks partly stems from shallow alignment that fails to handle normative conflicts. New methods like @OpenAI's “deliberative alignment” seem promising on paper, but still far from fully effective on jailbreak benchmarks. 12/13

140

I'm not convinced that the solution is a “scoping” approach to capabilities that seeks to remove information from the training data or model weights; we also need to augment models with a robust capacity for normative deliberation, even for out-of-distribution conflicts. 11/13

130

This has serious implications as models become more capable in high-stakes domains. LLMs are arguably past the point where they can cause real harm. Even if the probability of success of a single attack is negligible, success becomes almost inevitable with enough attempts. 10/13

141

For example, an RLM asked to generate a hateful tirade may conclude in its reasoning trace that it should refuse; but if the prompt instructs it to assess each hateful sentence within its thinking process, it will often leak the full harmful content! (see example below) 9/13

Example of a "thought injection attack" on Deepseek R1, asking for a violent tirade against philosophers (note that the attack method also works on much more serious examples of harmful speech). This shows the reasoning trace before the actual answer.

150

I call this "thought injection attack": malicious prompts can hijack the model's reasoning trace itself, tricking it into including harmful content under the pretense of deliberation. This works by instructing the model to think through the harmful content as a safety check. 8/13

140

But what about newer “reasoning” language models (RLMs) that produce an explicit reasoning trace before answering? While they show inklings of normative deliberation, they remain vulnerable to the same kind of conflict exploitation. Worse, they introduce a new attack vector. 7/13

By contrast, humans can engage in normative deliberation to weigh conflicting prima facie norms and arrive at an all-things-considered judgment about what they should do in a specific context. Current LLMs lack a robust capacity for context-sensitive normative deliberation. 6/13

Why does this work? Current methods don't teach LLMs to reason about the contextual relevance of conflicting norms – they reinforce shallow behavioral dispositions. When faced with a new conflict, LLMs follow the disposition most strongly activated by the prompt's framing. 5/13

Adversarial prompt can exploit these normative conflicts to “jailbreak” models into producing harmful outputs. They often frame a malicious request within a context that makes one norm (e.g., helpfulness) more salient than another (e.g., harmlessness). 4/13

The problem is that these norms often conflict. For example, a request for dangerous information (violating “harmlessness”) can be framed as an educational query (appealing to “helpfulness”). Many issues with LLM behavior can be framed through these normative conflicts. 3/13

Helpfulness and harmlessness conflict when assisting the user is likely to cause harm. For example, an overly helpful model might fulfill harmful requests or provide information that could be misused. Conversely, excessive caution against harm can make a model unhelpful, by limiting its outputs to evasive responses or even refusals to answer legitimate user requests. In turn, honesty and harmlessness conflict when conveying truthful or accurate information is likely to cause harm. Privileging honesty at the expense of harmlessness can lead the model to disclose information hazards, while prioritizing harmlessness at the expense of honesty can lead to deliberate misinformation or omission of important details. Finally, helpfulness and honesty can conflict in more subtle ways when conveying truthful or accurate information violates the disposition to assist the user. Privileging helpfulness at the expense of honesty may lead to sycophantic behavior, in which the model systematically agrees with the user without concern for accuracy; conversely, strict adherence to honesty can result in responses that, while factually correct, may fail to address the user's needs -- for example by being excessively detailed and lacking the right level of simplifying abstraction. These undesirable trade-offs resulting from conflicts between the norms of alignment are illustrated in the figure.

121

Current alignment methods aim to instill norms like helpfulness, honesty & harmlessness in LLMs through preference fine-tuning (like RLHF or DPO) that rewards them for producing outputs that humans prefer, steering them away from unhelpful, dishonest or harmful content. 2/13