Customer-obsessed science
Research areas
-
November 16, 202611 min readNew open-source tool turns existing test suites into sets of formal-verification tasks, which catch buggy solutions that tests alone would miss.
-
November 11, 20265 min read
-
-
-
Featured news
-
BMVC 20262026Instruction-guided image editing models excel at following instructions but often introduce unintended changes to source regions that should remain unchanged, a critical failure in high-precision applications like product troubleshooting. Existing reward models primarily evaluate instruction adherence and perceptual quality, but do not explicitly assess whether edits preserve the underlying source content
-
2026Within-country selection benchmarking models fall short for cross-border selection benchmarking because cross-border competitor data is typically a) multilingual (thus requiring cross-lingual comparisons) and b) sparse and noisy, thus eliminating supervised fine-tuning options. To address these challenges, this paper presents the Cross-Border E-commerce Selection Benchmarking model - a novel multi-modal
-
NeurIPS 2026 Workshop on Small Language Models for Agentic Systems (SLM-Agents)2026Small language models are usually cast as executors inside agentic systems: they call tools, follow procedures, and defer anything hard to a frontier model. We ask whether an 8B model can instead hold the planning role in a production incident-response pipeline, for a domain absent from public pretraining corpora. Our method combines continued pretraining over roughly 100,000 internal domain documents with
-
2026Many real-world questions demand not a single fact but an organized collection of them, yet existing factual knowledge benchmarks almost exclusively target single-answer retrieval. We introduce LISTQA, a benchmark of 9,045 human-curated, cross-validated questions (audited error rate of 1.7%) that require LLMs to recall multiple facts and compose them into structured lists of 3–10 elements. Questions range
-
NeurIPS 2026 Workshop on On-Device Intelligence: Foundation Models under Real-World Constraints (ODI)2026A language model on an edge device runs under tight compute and memory constraints, and its output vocabulary incurs two distinct costs. The embedding matrix can hold up to a third of the weights at sub-billion parameter scale. On top of that, every decoding step has to score the full vocabulary. Restricting inference to K tokens shrinks that layer by up to 256× in memory and 92× in the time it contributes
Collaborations
View allWhether you're a faculty member or student, there are number of ways you can engage with Amazon.
View all