The Archive

Hey all, Our ML team spent some time this week getting training and deployments working for Gemma-4, and wanted to document all the things we ran into along the way. * **PEFT doesn't recognize Gemma 4's custom layers.** Google wrapped vision/audio projections in a new `ClippableLinear` class that doesn't inherit from `nn.Linear`, so PEFT refuses to attach LoRA, even for text-only fine-tuning. Fix: unwrap the wrappers after loading weights but before calling PEFT. * **SFTTrainer killed training silently.** TRL hardcodes `use_cache=False`, which breaks Gemma 4's KV-sharing attention. Loss nev...

Benchmark of 10 frontier LLM agents on NYU CTF Bench offensive cybersecurity tasks across 7 providers, 200 challenges.

Lightweight VR cybersickness detection model using user-specific eye and head tracking data with low model complexity.

Bayesian and ensemble extensions of PINNs for uncertainty quantification in turbulent flow reconstruction from sparse data.

The company now offers robotaxi service in three cities, all of them in Texas, after launching in Austin last year and starting to offer rides without safety drivers in January 2026.

LLMs generate executable legal decision models from Dutch environmental law text using structured intermediate representations, evaluated on production Omgevingsloket platform.

FlowRefiner applies flow matching with ODE-based iterative refinement to improve autoregressive neural PDE solvers for 3D turbulent flow prediction.

Graph-of-Agents proposes graph-based multi-agent LLM orchestration with node sampling and decoupled communication to improve task performance over Mixture-of-Agents.

According to Nikkei Asia, even as suppliers ramp up DRAM production, manufacturers are only expected to meet 60 percent of demand by the end of 2027. SK Group chairman has even said that shortages could last until 2030. The world's largest memory makers - Samsung, SK Hynix, and Micron - are all working to add new fabrication capacity, but almost none of it will be online until at least 2027, if not 2028. SK opened a fab in Cheongju in February, but that is the only increase in production among the three for 2026. Nikkei says that production would need to increase by 12 percent a year in 2026 ...

Analyzes negative momentum convergence properties in convex-concave min-max optimization, extending prior work on divergence prevention.

SeekerGym benchmark evaluates completeness and trustworthiness of information retrieved by deep research AI agents, addressing information gaps and bias.

SciImpact introduces multi-dimensional benchmark for scientific impact prediction across 19 fields using citation, awards, media, patents, and adoption signals.

Local Inconsistency Resolution presents unified framework on Probabilistic Dependency Graphs that generalizes EM, belief propagation, and related inference algorithms.

Identifies vulnerability in majority-voting multi-agent LLM systems against prompt injection, proposes token-level collaboration for robustness.

BOIL applies PageRank and information maximization to extract environment structure insights for multi-agent navigation and coverage tasks.

RoIt-XMASA extends cross-lingual sentiment analysis dataset to Romanian and Italian with adversarial meta-learned training for domain-language invariance.

CGM-Agent enables privacy-preserving QA over continuous glucose monitor data using LLMs with local computation to protect sensitive health records.

Adaptive Contrastive Decoding mitigates over-refusal in safety-aligned LLMs by keeping non-refusal tokens in candidate lists while maintaining high refusal for malicious queries.

Neural network classifier identifies three plasma regions around Mars using ion energy spectra from MAVEN satellite data.

Cluster cleaning algorithm with logistic regression addresses PU classification under non-SCAR conditions.

CASCADE proposes three-tier defense architecture against prompt injection and tool poisoning attacks in Model Context Protocol-based LLM systems.

Analysis shows feedforward Transformers struggle with dynamic state tracking due to sequential dependencies, pushing state representations deeper in layer stack.

DenseNet201-UNet++ dual-stage framework segments ten gastrointestinal organs from coronal MR enterography for IBD diagnosis.

In recent months, the company announced an agreement with Amazon Web Services to use Cerebras chips in Amazon data centers, as well as a deal with OpenAI reportedly worth more than $10 billion.

Frontier LLMs fail citation accuracy for rare diseases; HEG-TKG grounds clinical claims in temporal knowledge graphs with evidence traceability to resolve provenance gap.

Cross-model semantic disagreement complements self-consistency for uncertainty quantification in LLMs, detecting overconfident errors when single-model consensus fails.

HIVEMIND applies OS scheduling primitives (admission control, AIMD backpressure, circuit breaking) to coordinate concurrent LLM coding agents sharing rate-limited API endpoints.

Despite recently being designated a supply-chain risk by the Pentagon, Anthropic is still talking to high-level members of the Trump administration.

New data from Appfigures shows a swell of new app launches in 2026, suggesting AI tools could be fueling a mobile software boom.

Anthropic's Claude system prompts published as git timeline with separate files per model version.