The Archive

CarryOn organizes long-horizon agent context around reusable task objects instead of interaction histories, reducing redundant processing.

BLADE traces failure attribution in LLM multi-agent systems by auditing where information went wrong, not just where it came from.

Jeva proposes structural requirements for System One decision-making models: null-awareness, permutation-invariance, option-scalability, and cacheability.

University of Michigan spinout Laplace Intelligence uses AI agents and interactive simulation to stress-test robots and self-driving vehicles before deployment. The profitable startup is competing at TechCrunch Disrupt's Startup Battlefield.

ReVeR derives rubric-based reward models from task instructions alone using tool-augmented judges to avoid response-induced bias.

LLM crisis reporting evaluation on five disasters shows Gemini 3.7-flash nominally above human ceiling but difference statistically insignificant.

Agentic code judging generates verifiable feedback without consuming official submission budgets by diagnosing failures in competitive programming.

PreciseCoder augments unit-test rewards with semantic verification signals to reduce hallucinated bugs and over-editing in RL-trained code models.

Pluto Security has raised $42 million Jump Capital, TLV Partners, Ansa Capital and angel investors.

SolBehave benchmarks behavioral correctness of LLM-generated Solidity contracts via differential fuzzing and transaction replay across 250 Ethereum contracts.

Gauge-equivariant GNNs exhibit loop geometry with rapid contour growth and slow spatial expansion independent of network training.

Rate-transfer method converts SDE Lyapunov estimates to SGD convergence bounds via stochastic calculus arguments.

Token-wise importance sampling for off-policy RLVR with KL regularization achieves unbiased policy gradients under coverage conditions without full-response weighting.

SP²ec adapts draft model configuration during inference to accelerate Vision-Language Model decoding via self-speculative decoding without retraining.

RL-based method selects when to transfer knowledge across manufacturing domains to avoid negative transfer and reduce data requirements for new processes.

Natural Language Autoencoders for model oversight must distinguish true activation explanations from false ones; discriminative oversight differs structurally from training.

Lightweight method detects code-switching in speech directly from pretrained foundation model embeddings without requiring transcription or multilingual ASR.

Transformers trained on sequences with context-dependent Markov sources learn source structure comparable to informed Bayesian predictors with ground-truth priors.

AIRRI framework generates machine-resistant educational content that prevents LLM/OCR cheating while maintaining human readability for academic integrity.

Facebook's new "Your Algorithm" feature will let you directly add and manage topics you want to see in your feed and Facebook Reels. Meta says your recommendations "respond right away" after you make updates using the new feature by hitting "interested" or "not interested" for different topics. Your Algorithm previously launched on Instagram earlier this year, and Meta also launched a similar "Dear Algo" feature on Threads. Facebook users will be able to access Your Algorithm or select "interested" or "not interested" from the three dots menu on individual posts. The "interested" and "not int...

Google plans to help turn those personal AI-generated apps into wide releases.

The next time you vibe-code a custom app with Google's AI Studio, you won't have to connect your phone to a computer to install it. On Tuesday, Google announced a new feature in the US that will allow you to launch personal apps directly through the Play Store. Google first launched the ability to vibe-code personal Android apps in May. Even though you can now launch your custom apps in Google Play, the company notes it will continue to keep them "secure, private, and separate from the public store catalog." You can also launch them using a standard Google account, and the company will soon a...

Google is making it easier to build and use personal Android apps created with AI, without having to make them available to everyone on the Play Store.

Non-clashing teaching dimension for planar graph concept classes is proved to be at most 2, advancing learning theory on teaching vs. learning complexity.

Steertuning enables robots to learn new manipulation tasks from 5 human videos via pretrained 400M-parameter diffusion policy on egocentric interaction data.

Training-free method attributes token-level rewards from process reward models by measuring token influence, enabling efficient fine-grained reward signal extraction.

UniW(A)M unifies world models and world-action models via mixture-of-experts to share dynamics while preserving distinct prediction and control capabilities.

Kronecker-preconditioned optimizer refresh placement in language model training exhibits learning-rate-dependent effects on held-out loss independent of refresh count.

GeoReader improves vision-language models' spatial reasoning by learning to actively use geometric representations aligned with latent features.