At Startup School 2026, we hosted YCML, our first machine learning research showcase.

The work presented there tackled a wide range of questions at the frontier of ML. Can small language models reason better without relying on a larger teacher model? Can a theorem prover keep learning new mathematics without forgetting what it already knows? How do you train a model to understand a complicated scientific chart, reason across a two-hour video, or make predictions directly from a relational database?

Here are eight of the papers presented, explained by the researchers behind them.

**1. Test-Time Scaling for Multistep Reasoning in Small Language Models via A* Search**

Alexander Braverman
[Paper]

Small language models are cheaper and faster to run, but they struggle more with complex reasoning. Improving them often means distilling knowledge from a larger teacher model or generating several possible answers and using a separate reward model to score them. Both approaches require additional models, training, or infrastructure.

Alexander Braverman’s method instead asks the small model to evaluate its own reasoning. It generates possible paths, scores them through self-critique, and uses an A*-inspired search algorithm to decide which branches to keep exploring, without relying on an external reward model.

Tested with Qwen3-4B on GSM8K and MATH-500, the method improved accuracy by three to four percentage points over other test-time scaling approaches at comparable token and runtime budgets.

That comparison matters. It is easy to improve test-time performance by simply spending more compute. The paper shows that the search itself is making better use of the same resources.

2. LeanAgent: Lifelong Learning for Formal Theorem Proving

Adarsh Kumarappan
[Paper]

Continual learning presents models with a difficult tradeoff. A model that is too stable can’t absorb new information; one that adapts too quickly may overwrite what it already knows.

Adarsh Kumarappan studied this problem through formal mathematics, where knowledge is cumulative and every answer can be checked. In Lean, proofs are written as code, and a compiler verifies whether each step is correct.

LeanAgent collects Lean repositories from GitHub, estimates theorem difficulty, and arranges them into a curriculum from basic to advanced. As it works through that curriculum, it updates a retriever that identifies useful premises, uses best-first tree search to construct proofs, and adds successful proofs back into its database.

LeanAgent produced 155 new formal proofs across 23 domains. It also showed backward transfer: after learning more advanced subjects, the system improved at areas it had studied earlier. Learning topology didn’t make it forget algebra. It made it better at algebra.

3. Diamond Maps: Efficient Reward Alignment via Stochastic Flow Maps

Douglas Chen
[Paper]

A generative model may produce excellent images but still be difficult to steer toward the specific image you want. To guide generation, a system must estimate whether an intermediate state is heading toward a high-reward result. The problem is that halfway through image generation, the output often still looks like noise.

Existing flow-map methods project that intermediate state into one possible final image and score it. But the same state could lead to many different outputs, making a single deterministic sample a weak estimate.

Douglas Chen’s Diamond Maps samples several possible outcomes instead. By evaluating multiple futures from the same intermediate state, it gets a better estimate of how promising that state actually is.

The researchers built both a fine-tuned version and a training-free version that attaches to an existing flow map at inference time. The latter allowed them to run text-to-image experiments with Flux without paying the considerable cost of fine-tuning the full model.

4. ChartNet: A Million-Scale, High-Quality Multimodal Dataset for Robust Chart Understanding

Jovana Kondic
[Paper]

Understanding a chart requires a model to read labels, interpret visual structure, compare values, and perform numerical reasoning across all of them. Real charts can also be far more complex than simple bar or line graphs, from dense scatter plots to figures in scientific papers.

Jovana Kondic and collaborators at MIT, IBM Research, and the MIT-IBM Watson AI Lab built ChartNet, a million-scale dataset and generation pipeline for chart understanding.

Their key insight is that charts are usually created programmatically. ChartNet starts with a seed image, recovers approximate plotting code with a vision-language model, modifies that code with a language model, and renders new charts with perfectly matched supporting data.

Each sample includes the image, plotting code, data table, natural-language summary, and question-answer reasoning traces. Fine-tuning on ChartNet improved every open-source model family tested, and a two-billion-parameter model trained with it outperformed GPT-4o on the team’s chart tasks.

IBM Research also used ChartNet in the training mixture for a new vision-language model that outperformed much larger models on chart and table extraction. The open-source dataset had already been downloaded more than 50,000 times when Kondic presented the work.

5. Relational Transformer: Toward Zero-Shot Foundation Models for Relational Data

Mark Žnidar
[Paper]

Predictive modeling on company data often begins by flattening a relational database into one large table. That makes the data easier for traditional models to consume, but it can discard relationships expressed through primary and foreign keys, which must then be reconstructed through feature engineering.

Mark Žnidar’s model treats the relational database itself as the input. It represents the database as a graph, samples the relevant neighborhood around the target entity, and serializes the resulting cells. Different encoders handle text, numbers, and other data types.

Specialized attention mechanisms then capture column distributions, the target entity’s own features, and information from connected rows and tables.

Despite having only 22 million parameters, the model scored 73 AUROC on the item-churn example presented in the talk. A four-billion-parameter language model scored 62.

The work points toward relational databases becoming a first-class input for predictive models rather than something data scientists must manually translate into a format the model understands.

6. SAGE: Training Smart Any-Horizon Agents for Long Video Reasoning with Reinforcement Learning

Jitesh Jain
[Paper]

People don’t watch every video in the same way. You might watch a ten-second clip straight through, but skim a Formula One race to find the half hour you missed. Jitesh Jain calls this any-horizon reasoning: adapting how much effort you spend to the length of the video and the question being asked.

Existing video agents struggle with long videos and open-ended questions. Many rely on temporal-grounding models that try to locate exactly when something happened, but those models remain unreliable, partly because long-video training data is scarce.

Jain’s system supplements visual search with transcripts and web search. Gemini 2.5 Flash generates synthetic question-answer pairs and tool-use trajectories, avoiding the high cost of having people annotate hours of video.

Reinforcement learning then teaches the agent when each tool is useful. Wrong answers after tool calls are penalized, while successful use of a visual tool earns an additional reward.

As videos became longer, the trained agent took more reasoning steps and its advantage over the base model grew, suggesting it had learned to adapt its effort to the horizon of the task.

7. On the Fine-Grained Planning Abilities of VLM Web Agents

Surgan Jandial
[Paper]

Web agents are usually graded on the final result: did the agent complete the purchase, find the right product, or successfully navigate the website? That reveals whether it succeeded, but not why.

Surgan Jandial studied the planning between seeing a webpage and choosing the next action. Evaluating those plans directly is expensive because each decision depends on the screenshots, text, and tool calls that came before it, along with what happened afterward. As models improve, their errors also become subtler and harder to spot.

Rather than grade every complete trace, Jandial breaks planning into individual capabilities. Can the agent put webpage states in the right order? Predict the next state? Choose between possible actions? Recognize and correct an error?

The tests repurpose existing datasets, so they require no new collection or human annotation. Most open-source models tested scored below 50 percent, even though the individual questions were relatively simple.

The fine-grained scores also correlated with performance on full web tasks. That makes them useful as a cheap diagnostic: for about $10, a researcher can identify which planning skills a model lacks before running a much more expensive end-to-end evaluation.

8. Mechanistic Interpretability for Steering Vision-Language-Action Models

Bear Häon
[Paper]

Interpretability becomes more urgent when a model can act in the physical world. Bear Häon’s Berkeley paper was the first interpretability work focused specifically on robot foundation models, asking whether techniques used to understand and steer language models could also make embodied systems safer.

A vision-language-action model contains transformer blocks with feed-forward networks. Haon’s team projected activity inside those networks onto the model’s embedding space, creating a rough dictionary of the concepts associated with different activations.

They grouped those representations into concepts such as fast, slow, high, low, or cautious, then amplified a group during the forward pass to steer how the robot carried out the same natural-language command.

Robot foundation models introduce risks that don’t exist when an AI system is confined to a screen. A language model might know how to do something dangerous; a sufficiently capable embodied system could potentially carry it out.

Haon has since founded the Physical AI Safety Institute, bringing together AI safety, robotics, and control theory. Its Science of Physical AI Safety workshop will examine what robot safety can borrow from those fields and how these systems should be evaluated.

—

Watch the full playlist here.

Root Access

Eight ML Papers, Explained by the Researchers…