Skip to content
Publications

Our Research

We study when an expensive or conventional AI computation can be replaced by an alternative—and what evidence is required to accept or reject that substitution.

Research philosophy

Evidence before substitution

Our research explores computational alternatives in AI, with an emphasis on checkable guarantees, declared evaluation criteria, and transparent negative results. We distinguish formally proven substitutions from empirical hypotheses and resource-limited outcomes.

We research efficient alternatives to conventional AI computation and demand explicit evidence before accepting the substitution.

Formal results supported by independently checkable evidence and explicit scope conditions.

PreprintJuly 2026

Proof-Carrying Optimality for Finite Identification under Bounded Adversarial Answer Errors

Vikram Lex

Zenodo · Under Review at JMLR

Develops proof-carrying methods for finite identification when answers may contain a bounded number of adversarial errors. Candidate construction is separated from independently checkable optimality certificates: all 30 primary cells receive two-sided proofs, while 300 of 303 declared sweep cells receive proofs and three are explicitly reported as resource-limit outcomes.

Scope: explicit finite behavior tables and declared probe alphabets. Verification is polynomial in the explicit table and supplied proof, not necessarily in a succinct game description.

Formal MethodsCertified SearchRobust Identification

Empirical Studies

Measured alternatives evaluated against declared quality, resource, and transfer criteria.

PreprintJuly 2026

Bounded-Horizon Local Transformer Training on CPUs: Quality, Throughput, and Memory

Vikram Lex

Research Square · Under Review at JAIR

Tests bounded-horizon local training as a CPU-constrained alternative to full-horizon Transformer training. Async RGC measured 1.382× BP throughput with higher memory use, but the prespecified held-out 1% non-inferiority criterion and TinyStories transfer criterion failed; readout equivalence was not established.

The throughput comparison used 29 stage-worker threads for RGC and 24 BP intra-op threads; measured PSS increased from 1.94 to 4.31 GiB.

Local LearningCPU TrainingNegative Results

Early Explorations

Preliminary directions reported with their present limitations and without product or advantage claims.

PreprintJune 2026

Quantum Hybrid Modules for AI: Attention, Optimization, and Verification on Near-Term Quantum Hardware

Vikram Lex

Quantum Machine Intelligence (Springer Nature) · Under Review

Explores quantum-classical modules for AI search, optimization, attention, and kernel methods. This is an early exploratory preprint; its current results do not establish a practical or end-to-end quantum advantage, and it does not represent a current product direction.

Quantum ComputingExploratory Research