<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Synthetic Data | Andrej Leban</title><link>https://andleb.netlify.app/tag/synthetic-data/</link><atom:link href="https://andleb.netlify.app/tag/synthetic-data/index.xml" rel="self" type="application/rss+xml"/><description>Synthetic Data</description><generator>Hugo Blox Builder (https://hugoblox.com)</generator><language>en-us</language><lastBuildDate>Mon, 16 Nov 2026 00:00:00 +0000</lastBuildDate><image><url>https://andleb.netlify.app/media/icon_hu0b7a4cb9992c9ac0e91bd28ffd38dd00_9727_512x512_fill_lanczos_center_3.png</url><title>Synthetic Data</title><link>https://andleb.netlify.app/tag/synthetic-data/</link></image><item><title>K2-Horizon: A Frontier Fleet of Open Source Language Models</title><link>https://andleb.netlify.app/publication/k2horizon/</link><pubDate>Mon, 16 Nov 2026 00:00:00 +0000</pubDate><guid>https://andleb.netlify.app/publication/k2horizon/</guid><description>&lt;p>This is the technical report for K2-Horizon, a fleet of six open-source language models from MBZUAI&amp;rsquo;s Institute of Foundation Models (IFM).&lt;/p>
&lt;p>My contribution, stemming from my internship in the summer of 2026, is Section 5.12.5: a multi-agent &lt;em>research community&lt;/em> simulation.&lt;/p>
&lt;p>It simulates a persistent, real-time world in which researcher agents (each with its own persona, discipline, and skillset) work on a shared pool of research tasks, both collaborating and competing. They coordinate over direct messages, group chats, file sharing, and shared git repositories; they buy and sell work and services on a marketplace for an in-game currency, and interact with a wider audience of hundreds of simulated public accounts corresponding to journalists, students, and aggregators on an X-like social network.&lt;/p>
&lt;p>They submit papers to a conference with a deterministic verifier and an LLM referee, compete on a leaderboard for hard open-ended problems, strive to gain the most influence on the social network, or simply amass the most currency by performing tasks for others. A second, collaborative configuration is targeted towards the hardest tasks: it removes the competition and is designed to study emergent collaborative patterns of many agents working toward a shared goal.&lt;/p>
&lt;p>As the simulation is non-blocking (the world moves on as the agents think), it results in very dynamic trajectories where, for example, agents have to pivot their research having been &amp;ldquo;scooped&amp;rdquo; by another, forming alliances, and much more. Such trajectories thus provide dynamic training data that cannot be obtained from single-agent trajectories and were used to train the K2-Horizon family of models.&lt;/p></description></item><item><title>CausalDS: Benchmarking Causal Reasoning in Data-Science Agents</title><link>https://andleb.netlify.app/publication/causalds/</link><pubDate>Thu, 09 Jul 2026 00:00:00 +0000</pubDate><guid>https://andleb.netlify.app/publication/causalds/</guid><description>&lt;p>The paper introduces a benchmark &lt;em>generator&lt;/em> for causal reasoning in agentic data-science workflows: every benchmark instance is synthetically generated rather than curated. In it, we:&lt;/p>
&lt;ul>
&lt;li>Generate fully synthetic &lt;em>scenes&lt;/em> pairing a sampled causal graph and structural causal model with tabular data and natural-language story;&lt;/li>
&lt;/ul>
&lt;ul>
&lt;li>Add an optional &lt;em>observation layer&lt;/em> that varies the data-analysis difficulty without changing the causal side;&lt;/li>
&lt;/ul>
&lt;ul>
&lt;li>Derive tasks spanning all three rungs of Pearl&amp;rsquo;s hierarchy which include non-answerable questions (i.e., testing abstention);&lt;/li>
&lt;/ul>
&lt;ul>
&lt;li>Make the benchmark fully parameterizable, so that it can be tailored to specific evaluation goals or grounded in empirical distributions obtained from real-world corpora.&lt;/li>
&lt;/ul>
&lt;p>We thus test the agents along five separate axes of &lt;em>agentic causal data science&lt;/em>: symbolic causal reasoning, data science, uncertainty quantification, abstention, and tool use/coding.&lt;/p>
&lt;figure id="figure-causalds-generation-and-evaluation">
&lt;div class="d-flex justify-content-center">
&lt;div class="w-100" >&lt;img alt="CausalDS generation and evaluation." srcset="
/publication/causalds/pipeline_hu556a92362509b572b68807d10bd0aed0_139855_b6220da6950582ff1001d287e26ff797.webp 400w,
/publication/causalds/pipeline_hu556a92362509b572b68807d10bd0aed0_139855_7b2c8b7ea2f9c17d33bdfe3423641b39.webp 760w,
/publication/causalds/pipeline_hu556a92362509b572b68807d10bd0aed0_139855_1200x1200_fit_q75_h2_lanczos_3.webp 1200w"
src="https://andleb.netlify.app/publication/causalds/pipeline_hu556a92362509b572b68807d10bd0aed0_139855_b6220da6950582ff1001d287e26ff797.webp"
width="760"
height="297"
loading="lazy" data-zoomable />&lt;/div>
&lt;/div>&lt;figcaption>
CausalDS generation and evaluation.
&lt;/figcaption>&lt;/figure>
&lt;figure id="figure-an-example-causalds-scene">
&lt;div class="d-flex justify-content-center">
&lt;div class="w-100" >&lt;img alt="An example CausalDS scene." srcset="
/publication/causalds/scene_example_hud9e74da28d0f09aaf66ff881cba7f841_215693_bbd507a2dc5ad34d7647738c45dcf9ca.webp 400w,
/publication/causalds/scene_example_hud9e74da28d0f09aaf66ff881cba7f841_215693_6b7676f79a5460010cf8351f83303225.webp 760w,
/publication/causalds/scene_example_hud9e74da28d0f09aaf66ff881cba7f841_215693_1200x1200_fit_q75_h2_lanczos_3.webp 1200w"
src="https://andleb.netlify.app/publication/causalds/scene_example_hud9e74da28d0f09aaf66ff881cba7f841_215693_bbd507a2dc5ad34d7647738c45dcf9ca.webp"
width="760"
height="401"
loading="lazy" data-zoomable />&lt;/div>
&lt;/div>&lt;figcaption>
An example CausalDS scene.
&lt;/figcaption>&lt;/figure>
&lt;p>We evaluate twelve contemporary agents — GPT-6 Astra, Claude Fable 5.1, Claude Opus 4.8, Gemini 3.1 Pro, and GPT-5.5, along with the open-weight GLM 5.3, DeepSeek V4.1 Flash, Kimi K3, K2-MoVA 36B-A4B, Qwen 3.6 35B, Kimi K2.6, and Gemma 4 26B — on a 100-task exam with realistically grounded composition and a 200-task hard exam, with three complete runs per model on each. GPT-6 Astra leads the realistic exam and GLM 5.3 leads the hard exam by CausalDSScore, and the evaluated capabilities dissociate: graph recovery is largely mastered across the field, while abstention, uncertainty quantification, and coding/tool-use/reasoning efficiency separate the models.&lt;/p>
&lt;p>The &lt;a href="https://huggingface.co/papers/2607.08093" target="_blank" rel="noopener">paper&lt;/a> was featured in &lt;a href="https://huggingface.co/papers/date/2026-07-10" target="_blank" rel="noopener">Hugging Face Daily Papers on July 10, 2026&lt;/a>.&lt;/p></description></item><item><title>New preprint: CausalDS</title><link>https://andleb.netlify.app/post/new-preprint-causalds/</link><pubDate>Thu, 09 Jul 2026 00:00:00 +0000</pubDate><guid>https://andleb.netlify.app/post/new-preprint-causalds/</guid><description>&lt;p>We recently submitted a new preprint titled &lt;a href="https://andleb.netlify.app/publication/causalds/">&lt;em>CausalDS: Benchmarking Causal Reasoning in Data-Science Agents&lt;/em>&lt;/a> to &lt;em>arXiv&lt;/em>.&lt;/p>
&lt;p>CausalDS is a benchmark &lt;em>generator&lt;/em> for causal reasoning in agentic data-science workflows: every benchmark instance is fully synthetically generated, and tests the agents along five separate axes of &lt;em>agentic causal data science&lt;/em>.&lt;/p>
&lt;p>Joint work with Yuekai Sun.&lt;/p></description></item></channel></rss>