Raphaël Millière

I’ll be at #ICML2026 in Seoul all week. If you’re interested in the intersection of cognitive science, mechanistic interpretability, and philosophy -- let’s chat! DMs are open. Check my website for more on what I work on (raphaelmilliere.com), or see below.

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

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

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

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

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

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

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

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

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

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

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

Despite extensive safety training, LLMs remain vulnerable to “jailbreaking” through adversarial prompts. Why does this vulnerability persist? In a new paper published in Philosophical Studies, I argue this is because current alignment methods are fundamentally shallow. 🧵 1/13

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

Despite extensive safety training, LLMs remain vulnerable to “jailbreaking” through adversarial prompts. Why does this vulnerability persist? In a new paper published in Philosophical Studies, I argue this is because current alignment methods are fundamentally shallow. 🧵 1/13

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

Despite extensive safety training, LLMs remain vulnerable to “jailbreaking” through adversarial prompts. Why does this vulnerability persist? In a new paper published in Philosophical Studies, I argue this is because current alignment methods are fundamentally shallow. 🧵 1/13

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

Despite extensive safety training, LLMs remain vulnerable to “jailbreaking” through adversarial prompts. Why does this vulnerability persist? In a new paper published in Philosophical Studies, I argue this is because current alignment methods are fundamentally shallow. 🧵 1/13

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

Despite extensive safety training, LLMs remain vulnerable to “jailbreaking” through adversarial prompts. Why does this vulnerability persist? In a new paper published in Philosophical Studies, I argue this is because current alignment methods are fundamentally shallow. 🧵 1/13

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

Despite extensive safety training, LLMs remain vulnerable to “jailbreaking” through adversarial prompts. Why does this vulnerability persist? In a new paper published in Philosophical Studies, I argue this is because current alignment methods are fundamentally shallow. 🧵 1/13

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