arXiv · paper ·June 2026
Evaluates five jailbreak-defense paradigms across seven instruction-tuned models (7B-31B) and five attack families, proving that prompt-time activation defenses are structurally blind to prefilling attacks. Introduces response-time probing on the first generated tokens (AUROC 0.97-1.00 across all models); composed with null-space steering, it drives prefilling attack success to zero with no false positives on benign inputs. Code, attacks, and per-sample results released.
arXiv · paper ·June 2026
An automated red-teaming study finding that LLM safety alignment does not always improve monotonically across model generations. Using quality-diversity evolution (MAP-Elites) to probe successive generations of an open-weight model family, it shows that a mid-generation release can be more vulnerable than both its predecessor and its successor, with evolved attack archives transferring unevenly across generations. These longitudinal patterns are invisible to static benchmarks and surface only through adaptive, longitudinal probing.
ICLR 2026 Workshop AIWILD · paper ·May 2026
A quality-diversity evolutionary framework operating at the semantic level for LLM adversarial testing. Uses MAP-Elites to maintain a diverse archive of interpretable attack strategies across behavioral dimensions, revealing distinct vulnerability profiles across GPT-4o-mini, Claude 3.5 Sonnet, and Gemini 2.0 Flash.
arXiv · paper ·April 2026
TritonMoE, a Mixture-of-Experts inference kernel written entirely in OpenAI Triton with no CUDA. A fused gate+up GEMM computes both SwiGLU projections from shared tile loads, eliminating 35% of global memory traffic. Reaches 89-131% of Megablocks throughput at inference batch sizes (up to 512 tokens) on both NVIDIA A100 and AMD MI300X, with full code portability across vendors.
Google, Technical Disclosure Commons · technical disclosure ·January 2026
This work addresses the tension between computational accuracy and operational speed in real-time risk systems. It proposes a multi-tier risk decisioning system managed by a predictive, cost-benefit routing engine.
Google, Technical Disclosure Commons · technical disclosure ·January 2026
This work addresses a key challenge in multi-tiered anomaly detection systems: balancing the operational cost of accurate analytical models with the accuracy of less expensive rule engines.
arXiv · paper ·January 2026
A memory architecture representing stored information as continuous fields governed by partial differential equations. Demonstrates 116% improvement on multi-session reasoning and 43.8% gains on temporal reasoning benchmarks, with multi-agent configurations achieving near-perfect collective performance through field coupling.
arXiv · paper ·January 2026
A distributed framework leveraging Apache Spark for statistically rigorous LLM evaluation at scale. Treats evaluation as a data-parallel problem, providing bootstrap confidence intervals, statistical significance tests (paired t-tests, McNemar's test, Wilcoxon signed-rank), and content-addressable response caching backed by Delta Lake. Reports linear scaling performance.
Google, Technical Disclosure Commons · technical disclosure ·November 2025
Framework combining formal verification, program synthesis, and machine learning to automatically generate verified distributed system implementations. Achieves 274x speedup with 60% latency reduction.
Google Cloud · whitepaper ·May 2025
Introduces ETLC (Extract, Transform, Load, Contextualize), adding semantic, relational, operational, environmental, and behavioral context to data pipelines for GenAI systems.
Google, Technical Disclosure Commons · technical disclosure ·April 2025
An innovative approach modeling context as interacting fields rather than discrete states, enabling natural context flow and dynamic evolution through partial differential equations.
Google, Technical Disclosure Commons · technical disclosure ·January 2025
An AI framework supporting complex debate scenarios and group decision-making through a tiered structure of language models that dynamically generate and evaluate arguments.
BITS Pilani · thesis ·December 2023
A comprehensive framework for enterprises to transform data assets into economic value, establishing methodologies for Data Monetization through a structured 12-month approach.
BITS Pilani / arXiv · paper ·December 2021
A protocol and platform based on Blockchain Technology that enables transparent processing of personal data throughout its lifecycle from capture, lineage to redaction.
ResearchGate · paper ·January 2021
A privacy-aware open protocol to prove without repudiation an entity's point-in-time presence, participation and location in physical or virtual space.
Zenodo · paper ·March 2016
A novel Cascading Algorithm leveraging Bayesian inference to facilitate arbitrage-free pricing of financial options through continuously refined price estimates.
Zenodo · paper ·April 2015
Coupled Temporal Deep Belief Networks with Conditional Restricted Boltzmann Machines to encode intra-market movements and temporary coupling with global indices.
Zenodo · paper ·July 2014
On the fly creation and calculation of Max-flow min-cut gap and approximation of multicuts with deep emulations of a fault-free mesh on a mesh with random faults.
Zenodo · paper ·May 2012
Algorithms for automatic on-the-fly wrapper creation for extractions from HTML and XML documents and compressing the adjacency list representation of graphs.
Amity University · paper ·November 2009
A discrete time-step optimization algorithm based on limited parameters of the sensor-target systems with OpenCV-based implementation.
Zenodo · paper ·February 2007
Wireless power transfer via resonant magnetic coupling, demonstrating power transfer efficiencies of 67% at a distance of 30 cm with applications to consumer electronics and medical implants.
No publications match your search.