Masafumi Oyamada小山田昌史（おやまだ まさふみ）

How did language models become thinking systems? This invited tutorial traces the ideas behind modern AI, focusing on test-time scaling, loops, harnesses, and recursive self-improvement (RSI). Drawing on examples from our research, it connects phenomena and methods, theory and interpretation, and practical applications.

Agent Behavior Maps place agents by how likely they are to take recorded actions in shared task histories. In ALFWorld, an agent’s success rate can be estimated from its evaluated neighbors, without collecting new environment runs for that agent.

Why can sampling more answers make Best-of-N less accurate? As sampling grows, success depends on whether correct or incorrect answers dominate the upper reward tail. Extreme value theory connects this competition to both accuracy gains and reward hacking.

Scarce target-language data changes the best pretraining recipe. The M^3 scaling law places model scale, data reuse, language mix, and stage structure on one loss surface, revealing a compute-dependent switch from monolingual single-stage to multilingual two-stage training.

End-to-end web-agent evaluation can take hundreds of hours. Minimal Failure Sets turn it into a lightweight coverage question: does a shortened HTML observation keep the elements that actually make the task succeed?

Under the same token budget, algorithm discovery improved more by spending tokens on stronger individual program edits than by generating more candidates. On Circle Packing, this suggests a quality-per-iteration bottleneck for evolutionary search with coding agents.

Browser agents usually know the page but not the organization behind the task. cotomi Act turns ordinary browsing into shared memory: task boards, timelines, and wiki knowledge that the agent can consult as future work unfolds.

Automated reviewers aligned best with human judgments when guided by official conference criteria, not by prompts that imitate reviewer behavior. The result suggests that peer-review automation needs carefully designed evaluation standards, while overly rigid rubrics can push LLM scores away from human assessments.

Web-agent context is not just a compression problem. The best page representation changes with model capability and thinking budget, so observation pipelines should adapt to the agent rather than standardize on one reduced format.

More generations can improve LLM answers, but majority voting has a limit. Best-of-infinity studies the N → ∞ regime to separate the accuracy a model could reach with unlimited sampling from the finite compute needed to get close.

Rich documents are hard for RAG because a page can mix charts, text, titles, and images. SCAN shows that retrieval improves when chunks follow semantic regions instead of tiny layout fragments, giving LLMs and VLMs the context they need without processing whole pages.

LLM inference scaling often depends on where a solution is split into steps. DISC makes that boundary adaptive: it uses rollout feedback to use coarser steps when progress is clear and finer steps when a decision looks risky, improving compute efficiency across math and coding benchmarks.

LaMDAgent reframes model adaptation as an iterative design loop: enumerate possible training actions, build a candidate model, evaluate it, and remember what worked. The key idea is that an LLM agent can search over complete post-training pipelines, not just tune isolated hyperparameters.

Model ancestry carries performance information. When LLMs are fine-tuned or merged, their descendants often inherit enough structure from their parents that a lineage-aware predictor can forecast benchmark behavior more accurately.

A small attribution model can become competitive when the training data has the right structure. SynQA shows that synthetic evidence paths can train a 1B-parameter model to identify answer-supporting context across several QA settings.

LLM judges are more reliable when they receive explicit evaluation criteria, not just a reference answer. For open-ended tasks, reliability is shaped by how scoring is specified, and that design choice can materially change results even with the same evaluator model.

LLM query expansion is not a free retrieval boost. It helps when the model already has enough knowledge about the query, but can hurt search quality when the query asks about unfamiliar concepts or under-specified intent.

Can reasoning data be mined from documents that were never written as step-by-step solutions? Reasoning CPT treats ordinary text as the visible trace of an author’s hidden thinking, reconstructs that process synthetically, and turns broad corpora into training data for reasoning.

A retrieved passage can make a model more or less certain even when the final answer is the same. This work treats output probability as a signal for how medical LLMs respond to evidence, revealing behavior that accuracy alone can miss.

The work asks whether a language model can invent useful model-improvement algorithms, not just apply human-written ones. In model merging, self-generated Python algorithms improve GSM8k from 70.1% to 76.1% and match the strongest MATH result among the compared human-designed methods.

Jellyfish shows that local 7B-13B LLMs can become strong data-preprocessing solvers when tuned on task descriptions, injected knowledge, and reasoning traces. The key lesson is practical: for cleaning tables, targeted instruction data can unlock large gains without relying only on bigger hosted models.

How should LLMs be used with tabular data? Table-aware LLM work can be organized around five practical routes: prompting, fine-tuning, RAG, agents, and vision-language models, making a fast-moving design space easier to compare and apply.

Broad summarization queries often need passages that do not match the query directly. LightPAL treats retrieval as navigation through a passage graph, using random walks to surface indirectly relevant context while improving retrieval and summarization metrics in many settings.

Tabular preprocessing asks whether LLMs can operate over structured records, not just fluent text. Across error detection, imputation, schema matching, and entity matching, the results are promising but uneven: LLMs can rival specialized tools on some benchmarks, while efficiency and reliability still shape their practical value.

Effective keyword augmentation for zero-shot text classification depends on more than semantic similarity. REDEX ranks candidate keywords by task-aware relevance, inter-class exclusivity, and intra-class diversity, improving performance across fully zero-shot and generalized zero-shot settings.

Can LLMs summarize better by working as an organization? This work models summarization as a DAG of cooperating LLM agents and shows a practical lesson: network-like and hierarchical workflows can improve faithfulness or quality in some settings, but the best structure depends on the task.

Can data preprocessing get LLM-style flexibility without sending sensitive tables to an external API? Jellyfish shows that instruction-tuned 7B-13B local models can handle multiple data cleaning tasks on modest hardware while keeping data closer to where it lives.

DeepJoin reframes joinable-table discovery as learned retrieval. Instead of comparing raw column values one by one, it trains a language model so joinable columns land near each other in vector space, then uses approximate nearest-neighbor search to find candidates quickly.

If “opera” in music is close to several movie genres, its influence should be distributed across them, not forced into one hand-made label match. ATP uses optimal transport to translate preference weight between attribute spaces, then compares users after that translation.

Retrieval-augmented QA depends not only on the passages supplied at inference time, but also on the quality of passages seen during training. FiD can overfit to that context quality, so a model trained on clean evidence may lose accuracy when retrieval becomes noisier.

Table augmentation is only valuable if it improves the model that will use it. The system makes enrichment an evaluable workflow: add external columns, train with the enriched table, and compare prediction quality before and after enrichment.

The useful signal for taxonomy enrichment is not only in the seed taxonomy. Musubu uses pretrained language models as a source of implicit parent-child knowledge, improving parent prediction when only a small number of hierarchy examples are available.

Entity matching breaks when two names look similar for the wrong reason, or different for a superficial one. The key idea is to measure how similarity changes after string transformations, so a gain or drop becomes evidence about whether two records name the same entity.

Can two accounts be matched when all you see is behavior, not demographics? Purchase logs and browsing logs can support likely identity matches when their histories are mapped into a shared behavioral representation.

Crowdsourced hierarchical labels often stop before the most specific class. Instead of discarding those incomplete answers, this method combines the label hierarchy with worker-specific reliability to improve leaf-label estimation.

Data lakes hide useful joins behind spelling variants, formatting differences, and semantic near-matches. PEXESO embeds textual table values as high-dimensional vectors, so join discovery can search by similarity in value space rather than exact string equality.

When people, posts, or devices move and change keywords, continuous spatial-keyword search can be maintained efficiently by treating each update as a targeted repair. The system identifies only the registered queries whose top-k answers may change, then refills those lists with grid-indexed spatial and keyword evidence.

When annotators choose Unsure, they may be flagging the hardest examples rather than adding unusable noise. This work treats those responses as boundary information, letting classifiers learn from ambiguity instead of throwing it away.

Expert feature engineering often hides in ordinary notebook code, but naming differences make those operations difficult to reuse. This work turns code into typed syntax patterns so recurring transformations can be discovered across data-analysis projects.

A column’s meaning is rarely contained in its header alone. Meimei uses cell values and neighboring columns as context, so a numeric field like height is not confused with age or temperature just because the values look plausible.

Feature engineering often asks the same range-aggregation question in slightly different ways. APA-tree caches partial aggregate results in an adaptive tree, so later queries can reuse prior work instead of rereading the same records from disk.

Sparse formats are efficient only when the data really stays sparse. CVS asks whether one compressed representation can handle both sparse and dense vector sets efficiently, while still supporting the operations needed by learning algorithms.

A single demographic predictor assumes the same behavioral signals matter for every customer. R-iSVM learns customer and item clusters from purchase histories, then trains a local expert for each customer group, using behavior as interpretable model structure rather than hand-built features.

Streaming systems often need mutable reference data, but ordinary continuous-query execution does not guarantee a consistent view of it. This work adapts database concurrency control to streaming: two-phase locking, snapshots, and optimistic validation preserve serializable results while processing continues.

High-rate data streams can stall when every tuple waits for its database transaction to finish. Decoupling transaction invocation from result waiting improved throughput by an order of magnitude in the reported experiments, while an order-preserving variant kept results aligned with the stream.