RIKEN AIP Workshop on LLM×ML 2026

Abstract to be announced.

In-context learning (ICL) is a fundamental calculation in modern large language models (LLMs), and elucidating its underlying principles may offer valuable insights into the foundations of computational and human language processing. Drawing on several works conducted over the past two years, this talk presents our efforts to understand ICL from an empirical, "physics-inspired" perspective. Specifically, by combining theoretical insights and experiments from synthetic experiments on small models with empirical observations in large-scaled Transformer models trained from natural language corpus, we propose a series of frameworks for understanding the computations underlying ICL in modern LLMs. Furthermore, we demonstrate how these insights can be leveraged to effectively improve ICL performance.

Generation of text and speech in natural languages can be modeled as a stochastic process. This idea dates back to the seminal work of Markov and, later, to that of Shannon and also underlies the recent development of large language models (LLMs). The stochastic processes corresponding to natural languages should be distinct from those that generate nonlinguistic sequences. One of the features that discriminate linguistic and nonlinguistic sequences is power-law behavior, which is universally observed across different languages. In statistical physics, such behavior suggests that natural languages are critical: They lie near a critical point separating two phases in a parameterized space of stochastic processes. However, testing this conjecture is not straightforward because natural languages do not have any controllable parameters. Here, we use LLMs as controllable effective models of natural languages. Through statistical analyses of texts generated by LLMs, we find that, when a parameter analogous to physical temperature is varied, LLMs undergo a phase transition. The transition separates a low-temperature phase with complex repetitive structures in generated texts from a high-temperature phase in which LLMs generate incomprehensible texts. At the critical point between these phases, generated texts display power-law behavior similar to that of natural languages and most closely resemble natural languages as measured by a standard metric in natural language processing. These findings strongly suggest that natural languages are indeed critical.

As AI systems become more capable, it is harder to get reliable evidence about how they behave. Benchmarks saturate quickly, evaluation signals can be exploited, and human experts cannot always verify the answers. In this talk, I present some of our recent work on these problems. First, I show how to estimate the Bayes accuracy, the highest accuracy any model can reach on a task. This helps us see when a benchmark is saturated. Next, I show how a deliberately designed limit on achievable performance can reveal benchmark contamination and test gaming by coding agents. The same limit can also be used in reward design to discourage cheating during training. Finally, I discuss scalable oversight with partial human expertise. Even when no single expert knows the correct answer, experts can reliably rule out options they understand. This partial feedback is enough to evaluate a model or agent's accuracy and can also be used to guide the search for better agent workflows.

Preference alignment has become a central tool to align large language models (LLMs) with human values and desired behavior, and its online variant can further advance their capabilities beyond what static data alone offer by repeatedly generating and labeling new responses with human feedback. But costly, time-consuming annotations limit its scalability, making sample efficiency essential, and existing methods typically seek to improve it by expanding coverage of the response space or selecting pairs with the highest uncertainty for comparison. Although they are effective, these methods quantify uncertainty from the perspective of the LLM, which may differ from the uncertainty perceived by the user: a pair can appear highly uncertain to the LLM even when past comparisons from the user already imply a clear ordering. In this paper, we propose multidimensional preference-guided online direct preference optimization (MPG-DPO), which leverages low-cost multidimensional evaluations from automated verifiers on the user historical comparisons to construct a personalized preference cone guiding response selection for online preference alignment. Specifically, it first discards responses that are confidently dominated under the estimated user preference, then selects two responses that are competitive under the retained preference cone yet exhibit the strongest unresolved disagreement. Theoretical analysis shows that, with high probability, the fraction of queries spent on cone-dominated responses vanishes as learning progresses. Experiments across multiple LLM families show that MPG-DPO outperforms strong contenders under matched query budgets.

Long-context inference allows large language models (LLMs) to use extensive documents, interaction histories, and other contextual information, but its memory and computational costs increase rapidly with context length. An emerging alternative is to encode context into temporary model-weight updates and remove the original context from the input. However, memorizing information in weights does not necessarily enable an LLM to recognize and use that information as context during subsequent inference. In this talk, I will introduce Addressable Context Encoding (ACE), a method designed to bridge this gap without task-specific meta-training. ACE associates each context chunk with a unique identifier during encoding, providing an explicit address through which the encoded information can later be accessed. A complementary prompting framework then specifies how the retrieved information should be selectively used as conversational context. Experiments across controlled context-use tasks and long-context benchmarks show that ACE substantially improves recall, continuation, retrieval, and the handling of conflicts between encoded information and pre-existing knowledge compared with conventional test-time context encoding. These results highlight a central challenge for context-to-weight methods: information must not only be stored in model parameters, but also remain identifiable, accessible, and usable as context.

This study investigates the internal information flow of large language models (LLMs) while performing chain-of-thought (CoT) style reasoning.Specifically, with a particular interest in the faithfulness of the CoT explanation to LLMs’ final answer, we explore (i) when the LLMs’ answer is (pre)determined, especially before the CoT begins or after, and (ii) how strongly the information from CoT specifically has a causal effect on the final answer.Our experiments with controlled arithmetic tasks reveal a systematic internal reasoning mechanism of LLMs.They have not derived an answer at the moment when input was fed into the model.Instead, they compute (sub-)answers while generating the reasoning chain on the fly.Therefore, the generated reasoning chains can be regarded as faithful reflections of the model’s internal computation.

Operational disaster response goes beyond damage assessment, requiring responders to integrate multi-sensor signals, reason over road networks, populations and key facilities, plan evacuations, and produce actionable reports. However, prior work largely isolates remote-sensing perception or evaluates generic tool use, leaving the end-to-end workflows of emergency operations underexplored. In this paper, we introduce Disaster Operational Response Agent benchmark (DORA), the first agentic benchmark for end-to-end disaster response: 515 expert authored tasks across 45 real-world disaster events spanning 10 types, paired with expert-verified, replayable gold trajectories totaling 3,500 tool-call steps. Tasks span five dimensions that cover the operational disaster-response pipeline: disaster perception, spatial relational analysis, disaster operational planning, temporal evolution reasoning, and multi-modal report synthesis. Agents compose calls from a 108-tool MCP library over heterogeneous geospatial data: optical, SAR, and multi-spectral imagery across single-, bi-, and multi-temporal sequences (0.015–10 m GSD), complemented by elevation and social vector layers. We comprehensively evaluate 13 frontier LLMs on our benchmark, revealing three persistent challenges: 1) disaster-domain grounding exposes unique failure modes (damage-semantic grounding, sensor-modality mismatch, and disaster-pipeline composition); 2) agents are doubly bottlenecked by tool selection and argument grounding, where gold toolorder hints improve accuracy by only 1.08–4.40%, and alternative scaffolds yield at most a 3.24% gain; 3) compositional fragility scales with trajectory length, the agent-to-gold gap widening from 7% to 56% on long pipelines. DORA establishes a rigorous testbed for operationally reliable disaster-response agents.

Muon has emerged as a strong optimizer for the matrix-valued parameters in large language model pretraining, approximately orthogonalizing its momentum with a few Newton–Schulz iterations. Existing theory either replaces this iteration with the exact polar factor it approximates, or treats it as a source of approximation error, and thus the iteration Muon actually runs can only hurt the guarantees. We show that finite Newton–Schulz can instead be beneficial for nonsmooth nonconvex optimization. To this end, we analyze Muon through the online-to-nonconvex conversion, which views the update rule of an optimizer as an online learner and converts its regret bound into a stationarity guarantee. We observe that the finite Newton--Schulz iteration smooths the discontinuous polar map into a Lipschitz map of the singular values, and Muon with finite Newton–Schulz can be regarded as an online learner with a smoothed spectral potential. Building on this observation, we prove that a Newton–Schulz depth growing only logarithmically in the target accuracy suffices for convergence to stationary points in nonsmooth nonconvex optimization, whereas Muon with the exact-polar update may fail to converge. The resulting sample complexity bounds match the best-known guarantees for nonsmooth nonconvex optimization and are optimal for smooth nonconvex optimization up to problem-dependent factors. The argument extends beyond Newton–Schulz to general spectral maps that are Lipschitz smoothings of the polar map.

Based on our recent work, I will discuss two functional aspects of Muon optimizer: it acts as a robust feature learner against anisotropic data (arXiv:2601.22652), while annihilating signal-unrelated noises (arXiv:2606.03899).

Abstract to be announced.