Amazon Science
banner
amazon.science
Amazon Science
@amazon.science
The latest news and research from Amazon's science community.

http://www.amazon.science
Passing benchmark tests isn't the same as being correct. BenchProofer uses automated reasoning to improve benchmarks, catching buggy solutions that tests alone would miss.
BenchProofer: Formally verified coding benchmarks, on demand
New open-source tool turns existing test suites into sets of formal-verification tasks, which catch buggy solutions that tests alone would miss.
www.amazon.science
November 18, 2026 at 5:03 PM
The SGLang workflow you already know now runs on AWS Trainium.

SGLang-Neuron: serve models and generate RL rollouts on Trainium — same server, same APIs, native PyTorch through TorchNeuron. #SGLangSummit
Deploy frontier AI models on AWS Trainium with SGLang-Neuron | Amazon Web Services
SGLang-Neuron is now publicly available through an AWS Deep Learning Container. With SGLang-Neuron, you can serve models and generate reinforcement learning rollouts with SGLang on AWS Trainium using ...
aws.amazon.com
November 12, 2026 at 10:07 PM
Introducing Mitra-v2, a tabular foundation model for both classification and regression. Pretrained entirely on synthetic tables, it leads the single-model field on the TabArena benchmark.
Mitra-v2: Scaling synthetic pretraining for tabular foundation models
Expanding the mix of synthetic data used to train a tabular foundation model enables it to adapt to new tables without new models and lead the single-model field on the TabArena benchmark.
www.amazon.science
November 11, 2026 at 12:40 AM
📣 Last call: Amazon Research Awards Fall 2026 proposals are due November 11. Successful applicants receive unrestricted funds, AWS promotional credits, and training resources.
Amazon Research Awards opens Fall 2026 call for proposals
Submissions open October 1 and close on November 11, 2026
www.amazon.science
November 7, 2026 at 12:49 AM
Amazon senior applied scientist Hamed Karimi received the 2026 ECML PKDD Test-of-Time Award for his 2016 paper on the Polyak–Łojasiewicz condition. #ECML2026
Hamed Karimi wins ECML PKDD Test-of-Time Award
Amazon senior applied scientist and coauthors showed that an overlooked mathematical condition from 1963 proves why frequently used machine-learning optimization methods work well.
www.amazon.science
November 6, 2026 at 7:49 PM
How do you keep 70-billion-parameter quality while running up to 20 times as fast? AWS researchers combined knowledge distillation with parallel decoding. For @bookingcom.bsky.social's Smart Filters, that meant responses 15 times as fast for 99% of queries.
Speeding up language models for travel applications
Two optimization techniques, when combined, can significantly improve a model’s speed and cost effectiveness without sacrificing quality.
www.amazon.science
November 4, 2026 at 1:08 AM
A text-to-speech system that knows which character is speaking, and voices them accordingly. Audible's controllable TTS profiles characters and attributes dialogue with 97% accuracy in English:
Better speech synthesis for audiobooks
“Controllable text to speech” makes speech more expressive, character voices more distinct and consistent.
www.amazon.science
November 3, 2026 at 3:21 AM
📢 The Amazon Research Awards Fall 2026 deadline has been extended to November 11.

Award recipients receive unrestricted funds, AWS promotional credits, and training resources. Submit your proposal today:
Amazon Research Awards opens Fall 2026 call for proposals
Submissions open October 1 and close on November 11, 2026
www.amazon.science
October 30, 2026 at 12:17 AM
Amazon's custom Trainium chip helps power the diffusion models behind Prime 360's cinematic highlights, reducing diffusion-model inference latency by more than 40%: https://amzn.to/4e1puZN
October 28, 2026 at 8:52 PM
TorchNeuron is now open source. Our new @pytorch.org backend translates code into operations executable on AWS Trainium with minimal or no code changes. Learn how the team used AI agents and hundreds of thousands of models to validate it.
The science behind TorchNeuron, Amazon’s new PyTorch backend for AWS Trainium chips
Runtime optimizations and a thorough evaluation process driven by agentic AI help ensure that models written in PyTorch will execute efficiently and reliably on Trainium systems.
www.amazon.science
October 23, 2026 at 4:47 PM
New AWS research introduces priming, a method that reduces pretraining costs for hybrid architectures by over 99% by leveraging pretrained Transformer weights, saving trillions of tokens and hundreds of thousands of GPU-hours.
Can we change an AI model's architecture without retraining it from scratch?
“Priming” reuses pretrained Transformer weights to build models with more efficient architectures at less than 1% of the original training cost and with throughputs up to twice as high.
www.amazon.science
October 23, 2026 at 12:11 AM
A great week with the language modeling community at @colmweb.org. Thanks to everyone who connected with our team. See you next year! #COLM2026
October 12, 2026 at 8:51 PM
🎤 Can you out-rhyme the AI agent? Stop by the Amazon booth at @colmweb.org for a live AI rap battle demo, plus talks on multi-agent systems and responsible AI for science: https://amzn.to/4jsMYKM #COLM2026
October 8, 2026 at 4:24 PM
AI-powered smart glasses that guide delivery drivers from van to the customer's doorstep, hands-free. Hear from Amazon applied scientist Yelin Kim on the computer vision and edge AI behind it at @colmweb.org. #COLM2026.
October 7, 2026 at 11:16 PM
Amazon Scholar and UT Austin professor @mattlease.bsky.social explores how to harness AI for scientific discovery while assessing potential risks.

Catch his Expo Talk at @colmweb.org Thursday at 1pm PT. #COLM2026
October 7, 2026 at 6:17 PM
Day 2 at @colmweb.org. Stop by the Amazon booth for talks on multi-agent orchestration, visual reasoning, and LLM agent evaluation: https://amzn.to/4jsMYKM #COLM2026
October 7, 2026 at 4:19 PM
Reposted by Amazon Science
Looking forward to presenting this talk on Responsible AI for Science on Thursday at the 3rd Annual Conference on Language Modeling (COLM)! @colmweb.org #colm2026
@amazon.science
@nsfsimonscosmicai.bsky.social
@texasscience.bsky.social
@utaustin.bsky.social
October 6, 2026 at 11:01 PM
Day 1 at @colmweb.org. We're here all week with live research talks, meet-the-scientist sessions, and conversations on what's next for language models. Stop by the Amazon booth. #COLM2026

More info: https://amzn.to/4jsMYKM
October 6, 2026 at 5:33 PM
"One workload is power-bound. The next workload is memory-bandwidth-bound. The next is memory-bound. It's one of the most interesting hardware design problems that we've seen in ages." Amazon SVP Peter DeSantis sat down with SemiAnalysis founder Dylan Patel at #AIInfraSummit:
Inside a decade of Amazon chip design, built workload by workload
At AI Infra Summit, Amazon SVP Peter DeSantis explained how real customer workloads guided every generation of Amazon's chips, from Nitro to Trainium.
www.aboutamazon.com
September 22, 2026 at 7:06 PM
Amazon Bio Discovery developed three AI approaches to accelerate antibody drug design: MochiBind (sequence-based affinity ranking), CA-MAP (developability prediction with batch effect correction), and an agent-guided design system with 46 lab-validated hits against a novel cancer target.
Advancing AI for biology: Teaching models to design and characterize antibodies
Three new papers from Amazon Bio Discovery address bottlenecks in AI-driven antibody engineering, from benchmarking binding predictors to experimentally validating de novo design.
www.amazon.science
September 21, 2026 at 7:27 PM
Years of iterating against the same benchmarks should, by textbook logic, produce overfitting. It largely doesn't.

New research explains why: strategies that generalize can be expressed in too compact a form to allow memorization, while the ones that overfit don't survive a compression.
Why don’t machine learning research agents overfit?
New research indicates that AI agents learn compressible models of data, which don’t have enough space to enable memorization.
www.amazon.science
September 10, 2026 at 8:01 PM
Reposted by Amazon Science
A blog post on some neat work with @zstevenwu.bsky.social and Martin Bertran: www.amazon.science/blog/why-don...
Why don’t machine learning research agents overfit?
New research indicates that AI agents learn compressible models of data, which don’t have enough space to enable memorization.
www.amazon.science
September 10, 2026 at 3:50 PM
When LLM judges agree, the right question is why. Shared prompts, model families, or training lineage can make a majority look stronger than it is. Dependence-aware aggregation via Ising models accounts for this, improving accuracy 9–14% over weighted majority vote.
When LLM judges agree, should we believe them?
Discounting the opinions of LLM judges with highly correlated outputs ensures that panels of judges reflect a true diversity of perspectives.
www.amazon.science
August 28, 2026 at 6:29 PM
How did a model upgrade make agents worse? By pairing real enterprise SOPs with functioning tools and ground-truth grading across 12 industries and 2,000+ tasks, SOP-Bench helps find such anomalies.
SOP-Bench: A new benchmark for evaluating AI agents on real business procedures
Extendable framework enables testing agents on the full set of capabilities required to successfully complete a procedure, not isolated proxy tasks.
www.amazon.science
August 25, 2026 at 3:59 PM
Amazon's Automated Reasoning Group started by demoing tools to prove AWS systems secure and correct.

A decade later, they have proved the Nitro Isolation Engine, cryptographic code, and S3 correct. Now they are applying the same techniques to AI.
A decade of mathematical certainty: Reflections on the Automated Reasoning Group
Ten years after we founded the Automated Reasoning Group, mathematical logic has moved from academic research into production services that secure millions of customer workloads — demonstrating that s...
www.amazon.science
August 11, 2026 at 4:33 PM