Kanishk Paul

Research

Open questions, measured carefully, reported honestly.

Independent research at the edge of translation, world-model agents, and post-transformer attention — each built from first principles on local hardware, with results reported straight, including the negative ones.

Research

Three lines of work, three different bets. What ties them together is method: falsifiable questions, instruments built rather than assumed, honest numbers, and a clear line between what is finished enough to show and what is held back until formal publication.

Machine translation / negative result — paper in prep

Register Obstruction

The three-way honorific register of Bengali and Hindi is grammatically forced but unmarked in English. This study measures how direct and English-pivot translation preserve it, then tests a preregistered contextual-obstruction predictor. The predictor did not survive robustness analysis; the register-loss measurement remains. Detailed analysis, method, and data stay private pending the paper.

Agents / world models — prototypes

ARC-AGI-3 World Models

Three independent agents that induce a game's hidden rules online, probe under uncertainty, and plan in an internal simulator. Rigorous offline evaluation (replay verifiers, held-out log pairs, baselines); modest live scores reported straight — best run 4/183 levels, aggregate ≈ 0.257% — with the exact broken invariants diagnosed rather than hidden.

Post-transformer attention — benchmark

ButterflyGate

A sub-quadratic O(n log n) structured replacement for self-attention, benchmarked forward-only on real Gemma-4-E2B and Llama-3.2-1B weights. Faster than dense attention past a ~1330-token crossover and monotonically further ahead out to 14k+ tokens. Reported as a speed/scaling result on an untrained gate — the efficiency characteristic, not modeling quality; mechanism held back pending write-up.

01

How I Work

  • Falsifiable, pre-registered predictions with fixed kill criteria.
  • Instruments built and validated before they are trusted.
  • Everything reproducible on a single 16 GB machine, no cloud.

02

On Honesty

Negative results are results: a zero live score with an exact root-cause diagnosis, or a fast mechanism whose untrained accuracy collapses, are reported as they are. Where ground-truth labels came from a model rather than a human, that is disclosed, because it changes what the numbers mean.

Priority note. The item-level analysis, unpublished method, and frozen dataset behind Register Obstruction are deliberately not in the public repositories, to preserve priority ahead of peer review. The public pages disclose the high-level negative verdict without releasing paper-critical machinery.

Get in touch about the research