Purusharth Malik

Currently tinkering with efficient RL and diffusion language models.

SmolLM2-360M's attention averaged over its heads. Use the arrow keys to move between words.

Research engineer at Conscious Engines in Bengaluru, where I work on reinforcement learning for both autoregressive and diffusion language models. Before this I led post-training of 30B models at Gnani.ai and built LLM systems at Swiggy, after studying mathematics at the University of Delhi and AI at CHRIST University.

My playlist, as CLIP sees it

Writing

Long-form notes from whatever rabbit hole I fell into that month. Posts sit near each other on the map when they cover related ideas. Pick a topic to filter.

  1. 15 min Automatic Speech Recognition (ASR) From Scratch (With Intuition) Why speech is so much harder than text for a transformer, and how Wav2Vec2, Whisper, Conformers and transducers get around it.
  2. 20 min Is Latent Reasoning The Future? Hierarchical and tiny recursive models reason in latent space instead of tokens. Is that where reasoning goes next?
  3. 7 min LoRA Can Now Make Your LLMs Multimodal! VoRA adds vision to a language model through LoRA layers instead of a separate encoder.
  4. 8 min You Don't Always Need An MCP Client What the Model Context Protocol actually is, and how to call an MCP server without a client.
  5. 14 min The Biggest Enemy of Generative Modeling The normalising constant, and how VAEs, GANs, score matching and diffusion each learned to avoid it.
  6. 16 min Is your LLM “playing along” with your guardrails? Alignment faking: when a model behaves well only because it believes it is being watched.
  7. 11 min How did we get to vLLM, and what was its genius? How LLM serving grew out of Flask apps, and why PagedAttention changed the economics of inference.
  8. 10 min Transformer²: The Death of LoRA? Sakana's Transformer² adapts weights at inference time with singular value fine-tuning. Does that make LoRA obsolete?
  9. 8 min Stop treating your LLM like a black box, you MORON! A first look at mechanistic interpretability: circuits, induction heads and reading what a transformer computes.
  10. 9 min Chain-of-Thought — All the Language Models are Just Extremely Intelligent Toddlers Why chain-of-thought works, seen through control theory and a very enthusiastic toddler.

Research

Reinforcement learning for language models: how to make it cheaper, and the odd ways it changes a model. Also on Google Scholar and alphaXiv.

  1. NeurIPS 2026 GlobalSouthAI Workshop

    arXiv 2611.16228 12 November 2026

    Indic-Task Gains and Regressions after Reinforcement Learning

    S. S. Raj, D. Maan, A. Benki, P. Chatterjee, P. Prasanjith, P. Malik

    We ran GRPO across four reward environments on a supervised 30B-A3B mixture-of-experts model covering English and eight Indian languages, then compared it before and after RL on 34 benchmarks. Maths jumped, but answers to Indic reading questions drifted into English, from 8.7% to 91.7% of responses.

    Abstract PDF Discuss on alphaXiv

    Score change after RL, in points
    MATH +36.07
    XQuAD English F1 +11.71
    IndicQA −8.64
    XQuAD Indic F1 −19.32

  2. NeurIPS 2026 Workshop on Foundations of LLM Post-Training in Changing Environments

    arXiv 2611.15717 11 November 2026

    Outside the Purview of the Reward: Capability Drift in Multilingual Post-Training

    S. S. Raj, P. Malik, P. Prasanjith, D. Maan, P. Chatterjee, A. Benki

    Post-training can improve what the reward measures while eroding what it never checks. Comparing two GRPO branches from one checkpoint, we find a language-aware reward recovers the lost Indic scores, reward-free distillation restores most of a later drop, and most of the measured MATH gain tracks the model simply boxing its answers more often.

    Abstract PDF Discuss on alphaXiv

    Two GRPO branches from one checkpoint
    MeasureReleasedIndic-aware
    MATH+36.07Preserved
    XQuAD Indic F1−19.32Recovered
    Answers in the requested scriptLostNot restored
    Competition maths and codeReferenceTraded off

  3. Technical report

    arXiv 2610.15371 11 October 2026

    A Sovereign, Indic-First, Reasoning Model

    S. S. Raj, P. Pasunuti, P. Chatterjee, D. Maan, P. Malik, N. Sarkar, A. Benki

    The technical report for Gnani-Evon-v3.3, a 30B hybrid Mamba-Transformer mixture-of-experts model for reasoning and tool use in English and eight Indian languages, built from NVIDIA's Nemotron-3-Nano with an expanded vocabulary, long-context continual pretraining and multi-environment RL.

    Abstract PDF Discuss on alphaXiv

    Change against the starting checkpoint, in points

    Released model

    Indic aggregate +5.2
    English retention −1.7

    Indic-aware branch

    Indic aggregate +14.6
    English retention −5.4

Work

Seven years plotted as a training run, with a drop at every new job or degree. Checkpoints are evenly spaced, so the years stretch where things moved fast. Select one to read the log.