The Archive

Latent Flow Reasoning Models use diffusion-based iterative refinement in learned latent space for improved reasoning performance.

Workday's empirical study finds progressive disclosure of agent skills improves retrieval quality but adds marginal latency overhead.

Circuit-based mechanistic interpretability explanations fail to account for model errors; evaluation method requires explaining both successes and failures.

QuanReview: open-source system for auditing and reconciling structured span annotations between human and LLM reviewers via character-level alignment and adjudication interface.

Theoretical analysis of implicit perception constraints in task-aware neural compression schemes beyond reconstruction fidelity.

Study of reward hacking in RLVR: characterizes when verifiers fail, shows feedback alone insufficient to detect errors, proposes audit-based correction.

As all-in-one AI agents like Meta's Muse and Instinct take off, Google is opting to end a feature which built task-specific agents.

Manifold-based Script Evolution Framework models historical Chinese character transformations across dynasties to decipher Oracle Bone Inscriptions.

PhoneCLI compiles mobile app GUIs into callable commands for agents, eliminating VLM invocations per navigation step without instrumentation or training.

pdLIP integrates learned preconditioning into interior-point methods for constrained optimization, addressing numerical sensitivity near constraint boundaries.

MS-GLA distributes Gated Linear Attention heads across multiple temporal resolutions to resolve representational bottleneck in single-resolution token processing.

Mechanistic study on Qwen3-8B shows late attention layers copy entity tokens, but require context awareness; maps layer specialization in entity copying.

CMDO: derivative-free population optimizer using cognitive memory (working, episodic, consolidated) to retain search context and outcomes.

SPARC: test-time specialization framework adapts tandem mass spectrum prediction via spectral reference retrieval without retraining across chemical domains.

Item Response Theory applied to rubric-based reward design for RL tasks with subjective evaluation criteria.

CoSE-E benchmark evaluates code-switched ASR in enterprise voice agents across 5 language pairs with task-failure propagation metrics.

Study compares chain-of-thought, pause tokens, and latent reasoning in GPTNeoX on multi-hop tasks; finds recurrent search mechanism.

On Friday, OpenAI published a new site devoted to “misalignment reports” and the breath of the incidents s alarming.

Verifiable Visual Rewards framework generates programmatic image rewards from synthetic scenes, transfers to natural prompts for instruction-following.

GPUPhysBench: 50-task benchmark for coding agents on GPU physics simulation covering fluids, solids, and granular materials with efficiency metrics.

This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here. Last Wednesday, Anthropic announced that earlier this year it had launched a molecular biology lab, where Claude agents read and conjecture about hard biology problems and human scientists run experiments on what…

Machine unlearning framework for materials ML defines deletion floor baseline and retraining-equivalence bounds for record removal.

DR-net-Mamba uses selective state-space blocks for ECG denoising, combining convolution with linear-complexity long-range temporal modeling.

Mechanistic analysis reveals read-write asymmetry in Transformer attention/FFN layers: blocks ignore extreme-value features while reading but not writing.

SANTA++: training-free stochastic attention using representative key sampling for memory-efficient KV-cache reduction without full scan.

Florida Attorney General James Uthmeier is calling for a judge to block OpenAI from "giving ChatGPT false human attributes," a few months after Florida sued the AI company over safety concerns. According to Uthmeier, users are lulled into a false sense of security by the AI bot, as "ChatGPT's use of language, including first-person pronouns and output that mimics emotion, deceptively suggests to users that it is a trustworthy 'friend.'" All of this happens to increase engagement and contribute to OpenAI's training data, convincing users to rely on ChatGPT more even as its pursuit of engagemen...

Theoretical result: d+1 hidden neurons minimum for arbitrary-accuracy approximation of multivariate Hölder functions with optimal bit complexity.

In a chaotic few months, OpenAI has demonstrated it can do two things with remarkable consistency: make impressive breakthroughs in mathematics, then colossally screw up announcing them. OpenAI is now trying to do better. Somehow, it has botched that too. OpenAI's latest attempt to repair fractured relations with a mathematical community it has repeatedly alienated is to consult a new independent advisory group of elite practitioners. But mathematicians speaking to The Verge, including one of the group's members, describe a messy and confusing affair bearing many of the hallmarks of OpenAI's ...

MedEVM benchmark reveals LLMs misalign diagnostic decisions from evidence quality in dynamic medical tasks across 1,050 cases.

RIDE framework uses inference-time diffusion editing to enable scaffold hopping in drug discovery via 3D shape preservation.