1 · Data Sources

2 · Theme Taxonomy

The taxonomy is a closed, manually-curated three-level hierarchy: 10 parent categories → 39 level-2 themes → 38 level-3 specialist themes (77 child themes total). Nothing is added automatically.

Raw tags that do not map to these 77 themes land in child_tag_classifications as child tags. They are visible in the theme table's expanded child rows. Promoting a child tag to a full taxonomy theme requires a human decision: edit config/__init__.py, add an OVERRIDE_MAPPING entry, append to taxonomy_changelog.md, re-run classify_tags.py --all --force.

Topic selection rationale

Every theme in the taxonomy was included because it represents a distinct, active research programme with its own benchmarks, venues, and hiring signals — not simply because it is popular. The bar for inclusion is that a theme must be specific enough to be trackable (i.e. a paper either is or is not about it) but broad enough to accumulate meaningful weekly counts. Themes that are too narrow become noise; themes that are too broad become meaningless aggregates. Below is the reasoning for each category and its child themes.

Foundation Models

These themes track architectural and capability choices at the base model level — the decisions that propagate downstream to every application. A shift here (dense → MoE, fixed → long-context) changes what every downstream product can do.

• Long Context — Context window length is the single most trackable proxy for capability expansion. The race from 4K → 128K → 1M tokens has been one of the clearest measurable trends in the field and directly determines which tasks (document analysis, multi-file coding, long conversations) are feasible at all.
• SSMs / Mamba — The first credible architectural alternative to transformers with sub-quadratic scaling. Tracked because if transformers are eventually displaced, state-space models are the leading candidate; a trend signal here has outsized strategic relevance.
• Mixture of Experts — The architecture behind GPT-4, Mixtral, and most frontier models' efficiency gains. MoE lets labs scale parameter count without proportional compute increases; understanding MoE adoption explains how the scaling laws are being extended.
• In-Context Learning — The emergent ability to adapt from examples in the prompt without weight updates. Foundational to understanding how models generalise and to evaluating few-shot versus fine-tuning tradeoffs.

Reinforcement Learning

RL is the primary mechanism for post-training — taking a capable base model and aligning it to be useful, safe, and correct. Almost every deployed frontier model has gone through some form of RL-based fine-tuning. The child themes track distinct stages of that pipeline.

• RLHF — Reinforcement Learning from Human Feedback is the technique that made ChatGPT viable; human preference labels train a reward model, which then fine-tunes the policy via PPO. Still the baseline alignment method and a prerequisite for understanding every variant that followed.
• RLAIF — Replaces expensive human labellers with AI-generated feedback (Constitutional AI, self-critique). Tracked separately from RLHF because it fundamentally changes the cost structure of alignment — enabling labs to scale feedback without proportional human labour.
• Online RL Fine-tuning — Running RL with live rollouts during training; the mechanism behind o1/o3-style reasoning models that learn to allocate thinking time before answering. This is the current frontier of post-training research.
• Process Reward Models — Reward models that score intermediate reasoning steps rather than just final outputs. Essential for making chain-of-thought training tractable at scale; directly coupled to the test-time compute trend.
• Model-Based RL — Learning an internal world model and planning within it rather than learning purely from environment interaction. Increasingly relevant as agents need to reason about multi-step consequences before acting.

Reasoning & Planning

Moving from pattern-matching to structured, multi-step reasoning is the defining capability challenge of the current generation. Benchmark performance on math, coding, and science correlates almost entirely with reasoning quality, making this category a proxy for overall progress.

• Chain-of-Thought — The prompting technique that first demonstrated that reasoning could emerge from scale. "Think step by step" shifted model capability in mathematics and logic overnight and remains the baseline for everything in this category.
• Test-Time Compute — Spending more compute at inference (repeated sampling, beam search, MCTS) to improve output quality. The o1/o3 paradigm; represents a fundamentally different scaling axis from training-time compute.
• Neuro-Symbolic — Combining learned representations with formal rules or logical constraints. Relevant for domains where reliability and verifiability are non-negotiable (formal proofs, legal reasoning, safety-critical decisions).
• Formal Verification — Proving correctness of model outputs using formal methods. Growing field at the intersection of programming languages research and AI safety; tracked because regulatory pressure will likely push high-stakes deployments toward verified outputs.

Agentic Systems

The shift from single-turn Q&A to multi-step autonomous agents is arguably the most commercially significant trend in applied AI. Every major lab is building toward agentic products. The child themes decompose the key sub-problems of agent architecture.

• Tool Use — Calling external APIs, running code, searching the web. What transforms a language model from a text generator into an agent capable of real-world actions. Without tool use, agents are bounded by their training data.
• Multi-Agent Coordination — Multiple specialised agents working in parallel or sequence. Enables division of labour and specialisation beyond what a single context window can handle; the architecture behind most enterprise agent frameworks.
• Memory & Retrieval — How agents maintain state and access relevant information beyond their context window limit. The core infrastructure challenge for long-running agents; without it, every session starts from scratch.
• World Models — Learned simulators of environment dynamics. Critical for planning agents that need to reason about consequences before acting, rather than reacting to each step in isolation.
• Computer Use — GUI and browser agents that operate software as a human would. The commercial frontier of agent deployment and the basis for most enterprise automation products; tracked because it is where agent research meets immediate commercial revenue.

Generative Models

Generative capability across modalities is both a research frontier and the basis for most consumer AI products. Tracked separately from foundation models because the research communities, benchmarks, and commercial dynamics are distinct.

• Vision-Language Models — Multimodal models that combine image understanding with language generation; now a standard capability in frontier models (GPT-4V, Gemini, Claude). Tracked because multimodal capability unlocks qualitatively new use cases.
• Video Understanding — Temporal reasoning over video frames. Relevant for content creation, surveillance, scientific analysis, and robotics; Sora demonstrated that video generation is within reach of current architectures.
• Audio & Speech — Recognition, synthesis, and generation. The interface layer for voice-first applications and a growing area of capability competition between labs and startups.
• Text-to-Image / Diffusion — Stable Diffusion, DALL-E, Midjourney. The consumer AI wave that demonstrated generative AI to the public; diffusion models are also spreading into non-image domains (protein structure, video, audio).

Efficient ML

Running capable models at acceptable cost and latency is the engineering constraint that determines which applications are commercially viable. Efficiency research often lags capability research by 12–18 months but has outsized economic impact once it arrives.

• Quantization & Compression — Reducing model precision (FP16 → INT8 → INT4) to fit models on consumer hardware without meaningful quality loss. The reason local AI deployment is possible and the key enabler of on-device inference.
• Speculative Decoding — Using a small draft model to predict tokens that a larger model verifies in parallel; delivers 2–4× inference speedup with no accuracy trade-off. Increasingly standard in production serving stacks.
• Custom Silicon — TPUs, H100/H200, Groq, Cerebras, Tenstorrent. The compute substrate beneath everything else. Tracked because silicon availability and pricing constraints determine who can build what, and the competitive dynamics here are moving fast.
• LLM Infrastructure — Serving frameworks, batching strategies, distributed inference, KV-cache management. The operational layer that bridges research models and production deployments at scale.
• Efficient Attention — FlashAttention, linear attention, sparse attention variants. Reducing the O(n²) memory and compute bottleneck that limits context length and throughput; a prerequisite for long-context models to be practically deployable.

Learning Methods

How models are trained — not just what data they see — determines how well capabilities generalise. These themes cover training-time innovations distinct from the RL post-training category.

• Synthetic Data — Using AI to generate training data at scale. Increasingly the only path to acquiring training signal in domains where human annotation is expensive, slow, or impossible. The strategic implication: labs that can generate high-quality synthetic data have a compounding data advantage.
• Continual Learning — Training models on new information without catastrophic forgetting of prior knowledge. Critical for deployed models that need to stay current without full retraining from scratch.

Interpretability & Safety

As models become more capable and more widely deployed, understanding what they are doing and ensuring they remain safe becomes both a research priority and a regulatory requirement. This category tracks the technical work on both fronts.

• Interpretability — Understanding which internal model features correspond to which concepts; mechanistic interpretability (circuits, sparse autoencoders, feature attribution) is the leading methodology. Tracked because interpretability is the prerequisite for principled safety work.
• Red-Teaming & Jailbreaks — Finding failure modes before deployment. The attack-defence dynamic drives both safety research and model hardening; each new jailbreak technique advances the field's understanding of where models are brittle.
• Alignment Theory — Constitutional AI, preference learning, value alignment — the formal frameworks for specifying what we want models to do and verifying that they do it. Includes DPO, reward modelling, and hallucination mitigation.
• Watermarking & AI Detection — Techniques for marking AI-generated content and detecting it at scale. Tracked because of growing regulatory relevance (EU AI Act, US executive orders) and because detection capability shapes how AI content is treated by platforms and institutions.

Scientific & Applied ML

AI's impact outside the AI field itself — in medicine, materials science, software engineering, and physical systems — is where much of the societal value will be realised. These themes track application domains with enough dedicated research activity to warrant their own signal.

• Code Generation — GitHub Copilot, Cursor, Devin, SWE-bench. Software engineering is the highest-value near-term application of AI; tracked as a leading indicator of AI's economic impact on knowledge work and as a proxy for model reasoning quality.
• AI for Science — AlphaFold, drug discovery, materials science, climate modelling. High-impact applications where AI is compressing decades of experimental research into months of compute. Distinct research community with its own venues (NeurIPS AI for Science workshop, Nature Machine Intelligence).
• Robotics & Embodied AI — Physical AI systems that perceive and act in the real world; Tesla Optimus, Figure, Boston Dynamics, and a growing set of academic labs. Tracked because embodied AI closes the loop from language to physical action, and commercial investment here is accelerating rapidly.

Evaluation & Benchmarking

Progress in AI is only meaningful if the measurement infrastructure keeps pace with capability. Benchmark saturation — models acing tests designed to be hard for humans — and benchmark contamination — training data leaking into test sets — make evaluation a first-class research problem rather than just a reporting exercise. This category tracks the work that determines whether any of the other trends in this dashboard are real.

3 · Labels & Tag Classification

Every paper ingested from arXiv is sent to Claude Sonnet, which reads the title and abstract and returns a list of free-form tags (e.g. lora-fine-tuning, sparse-autoencoders, pde-solving). These raw tags are the atomic unit of the pipeline — everything else (theme velocity, convergence scores, trends) is an aggregate of tag counts.

From raw tag → child tag → theme

Tag classification happens in three layers, applied in priority order:

Confidence levels

Each Claude-classified tag is assigned one of three confidence levels, stored in child_tag_classifications.confidence:

Flagged tags

During classification, Claude can flag a tag for human review — cases where it is genuinely uncertain between two different parent categories (not just between child themes). Flagged tags are written to child_tag_classifications.flagged = true. Run python pipelines/classify_tags.py --review locally to action them.

Where labels appear in the UI

Label registry & discovered labels

When classify_tags.py is run with the --approve flag on a tag, it is written to label_registry as an approved discovered label. These labels have a display_name, a parent_category (which may or may not match the preset taxonomy), and optionally a description. In the dashboard they appear in the Emerging ✦ sub-panel in the series sidebar and at the bottom of the theme table — never mixed into the 34 preset themes.

4 · Share of Total (chart Y-axis)

The main chart plots share of total, not raw paper count:

This normalises for the overall growth of arXiv — without it, every theme would trend upward simply because arXiv submissions have grown 3× since 2022. A theme at 8% means 8 out of every 100 papers that week touched that theme.

LOESS smoothing (bandwidth 0.30, adjustable in Chart Settings) is applied before drawing the line. LOESS (Locally Weighted Scatterplot Smoothing) fits a local polynomial at each point using a weighted window of neighbouring weeks. It removes week-to-week noise while preserving genuine trend direction.

5 · Convergence Score

A theme "converges" when multiple independent data sources spike simultaneously — suggesting a genuine breakthrough rather than a single noisy signal. The convergence score (0–4 scale) is pre-computed weekly in pipelines/convergence.py and stored in convergence_signals.

Spike detection

Each source maintains a 12-week rolling baseline (mean + standard deviation) per theme. A spike is declared when:

Score formula

Where spike_magnitude = (current − baseline_mean) / baseline_std. A score of 4.0 means all four sources spiked simultaneously at maximum weight. The score is normalised to 0–4 so it stays comparable across all historical weeks.

Signal strength labels

6 · Convergence Heatmap

The heatmap visualises the full history of convergence signals for all 34 themes simultaneously. Each cell is one theme × one week.

Rows are ordered by all-time paper count (highest-volume themes first within each category). The Fit all mode auto-sizes cells to your viewport width so the entire history fits without scrolling. Zoom in shows the top 20 themes at a larger cell size.

7 · Emerging Category Detection

The pipeline continuously scans child_tag_velocity for tags that may warrant promotion to the taxonomy. A tag is flagged as an emergence candidate only if it meets all five of the following criteria:

1. Volume ≥ 30 papers total — filters out one-off tags
2. ≥ 4 consecutive weeks of growth — sustained, not a single spike
3. Confidence is medium or low — not already confidently classified into an existing theme
4. First seen within 18 months — genuinely new, not a slowly-growing legacy tag
5. Not already top-3 child of its parent — must be truly breakout, not already dominant

8 · Why the chart may look different after a taxonomy change

When the taxonomy enforcement was tightened (May 2026), three things changed that affect chart appearance:

1. Active series reset. Previously, approved discovered labels (e.g. Sparse Autoencoders, Federated Learning) could appear in the series panel and be selected as active chart lines. Some of those had rising trends. After enforcement, only the 77 preset themes are charted — different set of lines, different trajectories.
2. LOESS right-edge bias. The most recent 4–6 weeks of any time series will appear to flatten or dip due to LOESS smoothing — this is always present, not a result of any data change. Disable LOESS in Chart Settings to see raw weekly values.
3. Partial current-week data. The most recent week's counts are incomplete until the pipeline's end-of-week run. The last data point is always understated.

The underlying share_of_total values stored in the database are not modified by taxonomy enforcement — only which themes are displayed changes.