Amazon Science

The latest news and research from Amazon's science community. #AmazonScience

[AWS Open Source @AWSOpen]

We spent tens of billions of tokens building a benchmark for AI security — so you don't have to.

Deception Benchmark is now open on GitHub: 14,822 samples, 16 languages, 70+ CWEs.

Download the dataset, run your tools, hold them accountable.

https://t.co/0wR2ELi8VG

#AWS #OpenSource

[The Warren Center for Network & Data Sciences @WarrenCntrPenn]

"Machine learning, at its core, is about generalization, not memorization," write @awscloud Applied Scientist Martin Bertran Lopez and @WarrenCntrPenn faculty affiliate @Aaroth for @AmazonScience. https://t.co/LVNEGVmMB1

[Amazon Science @AmazonScience]

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 bottleneck. https://t.co/R8vzj7B1s7

[Sasha Krassovsky @bztree]

Trainium has by far the best profiler of any accelerator. As you can see we have nanosecond-accurate traces of our programs, allowing you to write a program that generates images reliably in the profiler.

[Amazon Science @AmazonScience]

.@Amazon and @DARPA brought together 150+ researchers in Seattle last week, with speakers including @awscloud CEO @mattsgarman, Fields Medalist @TaoistTerence, @EPrinceton Associate Professor @BorisHanin, and @NSAGov's Michael O'Hara to explore how AI is transforming mathematical discovery and reasoning.

[Amazon Science @AmazonScience]

Amazon Redshift researchers were awarded Best Paper Runner-Up: Industrial Track at @VLDBconf for eliminating compilation cold starts in query execution – cutting compilation time from seconds to milliseconds with a 7x speedup on TPC-DS benchmarks. #VLDB2026 https://t.co/ErGuQN5bBL

[Amazon Science @AmazonScience]

Verus is an open-source, automated program verifier for Rust that mechanically checks code against a formal mathematical specification for all possible inputs. Amazon used it to prove correctness of Nitro Isolation Engine primitives. https://t.co/SuFeU5g8nf

[Amazon Science @AmazonScience]

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. https://t.co/7k6Bn4SjKd

[Matt Garman @mattsgarman]

Honored to be named to @TIME's 2026 TIME100 AI list. This one belongs to our customers and the teams at @awscloud. Still early.

[Amazon Science @AmazonScience]

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. https://t.co/B2NarMII2B

[Amazon Science @AmazonScience]

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. https://t.co/TilPN62pRQ

[Amazon Science @AmazonScience]

📣 AWS Trainium Frontier is open for registration. Train language models from scratch on purpose-built AI chips for @NeurIPSConf.

Prizes include $25K for first place, co-publication with Annapurna Labs researchers, and a presentation in Sydney. Deadline is September 30. #NeurIPS2026 https://t.co/SUw1WCy4iQ

[Amazon Science @AmazonScience]

Attending #KDD2026? Come meet Amazonians and fellow attendees at our mixer. Spots are limited – request to attend: https://t.co/eI4ggc6wDM

[Amazon Science @AmazonScience]

📣 Amazon Research Awards announces the 34 recipients of the Build on Trainium program, a $110M credit initiative supporting AI research at 30 universities on AWS Trainium: https://t.co/Xz3ZsCFkSn

[Amazon Science @AmazonScience]

Training a graph neural network on multiple objectives usually means blending conflicting gradients at every step. Instead of compromising among parameter updates from different training objectives, ControlG allocates capacity to objectives sequentially and dynamically via PID control. #ICML2026 https://t.co/wAwi6x0jn2

[Amazon Science @AmazonScience]

Most healthcare AI benchmarks test static medical knowledge or evaluate tool-using agents on provider-facing tasks. PatientAgentBench generates synthetic patient records and clinical vignettes, then runs multiturn dual-agent conversations scored by an LLM-as-a-jury panel across over 100 clinician-vetted criteria. https://t.co/CXovl0aZ6j

[Amazon Science @AmazonScience]

As AI agents take on higher-stakes decisions, their actions need to be provably correct. Amazon is investing in the Lean FRO to make mathematical proof accessible to every developer: https://t.co/2t73B4rh8j

[Amazon Science @AmazonScience]

@ZeeshanZiaML Welcome, Zeeshan!

[Amazon Science @AmazonScience]

https://t.co/4DIoInCbp0

[Amazon Science @AmazonScience]

Turnstile is an open-source Rust proxy that records token-native rollout data for agentic RL. Harness stays unchanged. Trainer gets exact token IDs, log probs, and loss masks. https://t.co/x4O1jwx5cg