AI Alignment

Opening — Why this matters now Multimodal large language models (MLLMs) are everywhere: vision-language assistants, document analyzers, agents that claim to see, read, and reason simultaneously. Yet anyone who has deployed them seriously knows an awkward truth: they often say confident nonsense, especially when images are involved. The paper behind this article tackles an uncomfortable but fundamental question: what if the problem isn’t lack of data or scale—but a mismatch between how models generate answers and how they understand them? The proposed fix is surprisingly philosophical: let the model contradict itself, on purpose. ...

Opening — Why this matters now Multimodal models can answer visual questions with alarming confidence. They can also be catastrophically wrong while sounding perfectly reasonable. The uncomfortable truth is that many vision–language models succeed without actually seeing what matters. They talk first. They look later—if at all. The paper behind LaViT puts a name to this failure mode: the Perception Gap. It is the gap between saying the right thing and looking at the right evidence. And once you see it quantified, it becomes hard to ignore. ...

Opening — Why this matters now Multimodal large language models (MLLMs) are getting better at seeing—but not necessarily at knowing. Despite steady architectural progress, hallucinations remain stubbornly common: models confidently describe objects that do not exist, infer relationships never shown, and fabricate visual details with unsettling fluency. The industry response has been predictable: more preference data, more alignment, more optimization. ...

Opening — Why this matters now LLM agents today are voracious readers and remarkably poor conversationalists in the epistemic sense. They browse, retrieve, summarize, and reason—yet almost never talk back to the knowledge ecosystem they depend on. This paper names the cost of that silence with refreshing precision: epistemic asymmetry. Agents consume knowledge, but do not reciprocate, verify, or negotiate truth with the world. ...

Why this matters now Reinforcement Learning from Human Feedback (RLHF) has become the de facto standard for aligning large language models with human values. Yet, the process remains painfully inefficient—annotators evaluate thousands of pairs, most of which offer little new information. As AI models scale, so does the human cost. The question is no longer can we align models, but can we afford to keep doing it this way? A recent paper from Politecnico di Milano proposes a pragmatic answer: inject Bayesian intelligence into the feedback loop. Their hybrid framework—Bayesian RLHF—blends the scalability of neural reinforcement learning with the data thriftiness of Bayesian optimization. The result: smarter questions, faster convergence, and fewer wasted clicks. ...

Personalization is the love language of AI. But today’s large language models (LLMs) are more like well-meaning pen pals than mind-reading confidants. They remember your name, maybe your writing style — but the moment the context shifts, they stumble. The CUPID benchmark, introduced in a recent COLM 2025 paper, shows just how wide the gap still is between knowing the user and understanding them in context. Beyond Global Preferences: The Rise of Contextual Alignment Most LLMs that claim to be “personalized” assume you have stable, monolithic preferences. If you like bullet points, they’ll always give you bullet points. If you once asked for formal tone, they’ll keep things stiff forever. ...

Vision-language models (VLMs) may describe what they see, but do they truly understand what they’re looking at — especially in social contexts? A recent paper introduces Cognitive Chain-of-Thought (CoCoT), a deceptively simple yet remarkably effective prompting strategy that helps these models reason like humans: through layered cognition, not flat logic. The Problem with Flat Reasoning Traditional Chain-of-Thought (CoT) prompting, while powerful for math and symbolic tasks, falls short when it comes to social or moral interpretation. Consider a scene where a person wears a mask indoors, and another says, “Hiding from the paparazzi, huh?” CoT may recognize the mask, but often misfires in guessing intent — is it a joke? A warning? An instruction? ...

When a large language model (LLM) answers your question with a high degree of confidence, do you trust it? What if it’s wrong—but still confident? The stakes are high in real-world applications, from legal guidance to enterprise decision support. Yet today’s LLMs remain notoriously unreliable in aligning their confidence with correctness. The paper Deliberative Searcher: Improving LLM Reliability via Reinforcement Learning with Constraints (Yin et al., 2025) offers a bold response: rewire LLMs to be reasoning-primary and information-secondary. Instead of front-loading search and passively absorbing evidence, Deliberative Searcher acts more like a prudent investigator: it thinks, self-assesses, retrieves external information only when needed, and calibrates its confidence step-by-step. Crucially, it learns this behavior through a custom constrained reinforcement learning regime. ...

What if your AI model isn’t just answering questions, but living in its own version of time? A new paper titled The Other Mind makes a bold claim: large language models (LLMs) exhibit temporal cognition that mirrors how humans perceive time — not through raw numbers, but as a subjective, compressed mental landscape. Using a cognitive science task known as similarity judgment, the researchers asked 12 LLMs, from GPT-4o to Qwen2.5-72B, to rate how similar two years (like 1972 and 1992) felt. The results were startling: instead of linear comparisons, larger models automatically centered their judgment around a reference year — typically close to 2025 — and applied a logarithmic perception of time. In other words, just like us, they feel that 2020 and 2030 are more similar than 1520 and 1530. ...

What makes a large language model (LLM) biased? Is it the instruction tuning data, the randomness of training, or something more deeply embedded? A new paper from Itzhak, Belinkov, and Stanovsky, presented at COLM 2025, delivers a clear verdict: pretraining is the primary source of cognitive biases in LLMs. The implications of this are far-reaching — and perhaps more uncomfortable than many developers would like to admit. The Setup: Two Steps, One Core Question The authors dissected the origins of 32 cognitive biases in LLMs using a controlled two-step causal framework: ...