<?xml version="1.0" encoding="utf-8"?>
<feed xmlns="http://www.w3.org/2005/Atom">
  <title>Aditya Lab: research and notes</title>
  <id>https://adityalab.in/</id>
  <link rel="self" href="https://adityalab.in/feed.xml" />
  <link href="https://adityalab.in/research" />
  <updated>2026-11-05T00:00:00+05:30</updated>
  <author><name>Aditya Lab</name></author>
  <entry>
    <title>Reading 22 Indian scripts with a small model</title>
    <id>https://adityalab.in/research/ocr</id>
    <link href="https://adityalab.in/research/ocr" />
    <published>2026-11-05T00:00:00+05:30</published>
    <updated>2026-11-05T00:00:00+05:30</updated>
    <author><name>Aditya Lab</name></author>
    <summary>How a 0.9B open vision-language model went from 36 to 76 word accuracy on Sarvam's Indic OCR bench across 23 languages, and why long pages still end too soon.</summary>
  </entry>
  <entry>
    <title>Vedic v1: speech to text in Indian languages, and what the score hides</title>
    <id>https://adityalab.in/research/stt</id>
    <link href="https://adityalab.in/research/stt" />
    <published>2026-11-05T00:00:00+05:30</published>
    <updated>2026-11-05T00:00:00+05:30</updated>
    <author><name>Aditya Lab</name></author>
    <summary>How Aditya Lab turned an open speech model into a no-settings tool for 22 Indian languages and English, and why word error rate misleads in these languages.</summary>
  </entry>
  <entry>
    <title>The two words our OCR model got wrong were right</title>
    <id>https://adityalab.in/blog/the-two-words-our-ocr-model-got-wrong</id>
    <link href="https://adityalab.in/blog/the-two-words-our-ocr-model-got-wrong" />
    <published>2026-11-05T00:00:00+05:30</published>
    <updated>2026-11-05T00:00:00+05:30</updated>
    <author><name>Shivam Vaghani</name></author>
    <summary>Our reading model disagreed with the volunteers' proofread text twice on a held-out Gujarati page. Both times, the 1890s print agreed with the model.</summary>
  </entry>
  <entry>
    <title>NIF: when should useful work become reusable capability?</title>
    <id>https://adityalab.in/research/nif</id>
    <link href="https://adityalab.in/research/nif" />
    <published>2026-09-28T00:00:00+05:30</published>
    <updated>2026-10-20T00:00:00+05:30</updated>
    <author><name>Aditya Lab</name></author>
    <summary>NIF investigates reusable capability with composition, verification and reversibility. Read the architecture intent and current implementation limits.</summary>
  </entry>
  <entry>
    <title>Reusing a verified answer in chat</title>
    <id>https://adityalab.in/research/nif-chat</id>
    <link href="https://adityalab.in/research/nif-chat" />
    <published>2026-10-20T00:00:00+05:30</published>
    <updated>2026-10-20T00:00:00+05:30</updated>
    <author><name>Aditya Lab</name></author>
    <summary>When a maths question returns in another language, a verified earlier answer matches thinking mode at about a fifth of its tokens, with a meaning check.</summary>
  </entry>
  <entry>
    <title>A faster 4-bit layer for our own adapters</title>
    <id>https://adityalab.in/research/nf4-kernel</id>
    <link href="https://adityalab.in/research/nf4-kernel" />
    <published>2026-10-20T00:00:00+05:30</published>
    <updated>2026-10-20T00:00:00+05:30</updated>
    <author><name>Aditya Lab</name></author>
    <summary>Where a 24B model spends its time while it writes, a hypothesis that turned out wrong, and a measured 1.15x to 1.43x server speed-up with exact 4-bit weights.</summary>
  </entry>
  <entry>
    <title>Aditya v12 evaluation</title>
    <id>https://adityalab.in/research/v12</id>
    <link href="https://adityalab.in/research/v12" />
    <published>2026-09-29T00:00:00+05:30</published>
    <updated>2026-10-04T00:00:00+05:30</updated>
    <author><name>Aditya Lab</name></author>
    <summary>Saved Aditya v12 evaluations in Indian languages: maths, knowledge, refusal and over-refusal beside the sarvam-m base model, with methods and limitations.</summary>
  </entry>
  <entry>
    <title>Stage R v8: reasoning gains across native and Romanized Indic</title>
    <id>https://adityalab.in/research/stage-r-v8</id>
    <link href="https://adityalab.in/research/stage-r-v8" />
    <published>2026-09-28T00:00:00+05:30</published>
    <updated>2026-09-29T00:00:00+05:30</updated>
    <author><name>Aditya Lab</name></author>
    <summary>Recorded GSM8K-Indic and MMLU-Indic results for Sarvam-M and Stage R v8, with real language slices, corrected baseline, regression and methodology limits.</summary>
  </entry>
  <entry>
    <title>From a model proposal to a checked operation</title>
    <id>https://adityalab.in/research/architecture</id>
    <link href="https://adityalab.in/research/architecture" />
    <published>2026-09-28T00:00:00+05:30</published>
    <updated>2026-09-29T00:00:00+05:30</updated>
    <author><name>Aditya Lab</name></author>
    <summary>How Aditya separates model generation, task state, project tools, permission policy and verification in a visible agent run, and what each part records.</summary>
  </entry>
  <entry>
    <title>Models, RAG, MCP and NIF answer different questions</title>
    <id>https://adityalab.in/research/compare</id>
    <link href="https://adityalab.in/research/compare" />
    <published>2026-09-28T00:00:00+05:30</published>
    <updated>2026-09-29T00:00:00+05:30</updated>
    <author><name>Aditya Lab</name></author>
    <summary>Understand the different roles of a foundation model, retrieval, a tool protocol and NIF. An architectural comparison, not a performance leaderboard.</summary>
  </entry>
  <entry>
    <title>What is Neural Intelligence Fabric?</title>
    <id>https://adityalab.in/blog/what-is-neural-intelligence-fabric</id>
    <link href="https://adityalab.in/blog/what-is-neural-intelligence-fabric" />
    <published>2026-07-18T00:00:00+05:30</published>
    <updated>2026-09-28T00:00:00+05:30</updated>
    <author><name>Shivam Vaghani</name></author>
    <summary>A direct explanation of Neural Intelligence Fabric: what it is, its capability lifecycle, its safety model, and what is implemented today versus proposed.</summary>
  </entry>
  <entry>
    <title>Why tool calling is not enough</title>
    <id>https://adityalab.in/blog/why-tool-calling-is-not-enough</id>
    <link href="https://adityalab.in/blog/why-tool-calling-is-not-enough" />
    <published>2026-07-18T00:00:00+05:30</published>
    <updated>2026-09-28T00:00:00+05:30</updated>
    <author><name>Shivam Vaghani</name></author>
    <summary>Why emitting a valid function call is not the same as finishing real engineering work, and what an agent needs between the call and a checked result.</summary>
  </entry>
  <entry>
    <title>Building coding agents that can prove the work</title>
    <id>https://adityalab.in/blog/building-verifiable-coding-agents</id>
    <link href="https://adityalab.in/blog/building-verifiable-coding-agents" />
    <published>2026-07-18T00:00:00+05:30</published>
    <updated>2026-09-28T00:00:00+05:30</updated>
    <author><name>Shivam Vaghani</name></author>
    <summary>A practical evidence model for coding agents you can trust: what to record, which checks can disagree with the agent, and when a task counts as done.</summary>
  </entry>
</feed>
