The Archive

First federated learning study on SpeechLLM-based end-to-end ASR with communication-efficient optimization for distributed training.

Part 3 of a series where we take public “token saver” add-ons for coding agents and run the same paired A/B benchmark against each of them. Part 1 was the caveman skill (advertised −65%, measured −8.5%). Part 2 was rtk (advertised −60–90%, measured +7.6%). We ran 80 paired tasks to test the ponytail skill for […]

Google DeepMind releases Gemini Robotics 2, extending multimodal foundation model capabilities to whole-body robot control and coordination.

Recursive’s $400 million outlay represents the bulk of the company’s fundraising to date.

Deep learning models reconstruct urban air pollution fields from sparse monitoring stations in Paris.

Google Search AI Mode assists with offline activities like concert ticketing and event planning.

Google Search AI features generate menu and tablescape suggestions for dinner party planning.

Cognivia LLM-based copilot for Cognitive Behavioral Therapy with domain-specific cognitive distortion annotations.

Rashomon Alignment measures functional model similarity geometrically across input space, beyond data-distribution-dependent metrics.

DynaBridge multimodal framework fuses acoustic, visual, textual data with LLM-generated summaries for DASS mental health assessment.

DecoEvo decouples scoring to co-evolve LLM solver and rubric generator in text space, addressing evaluation bottlenecks.

Controlled study compares LLM (ChatGPT-4o, Gemini) vs. human literature review performance across physics, astrophysics, cosmology.

OmniDelta allocates token budgets by skill across modalities in OmniLLMs, compressing audio-video sequences for inference efficiency.

Physics-informed neural operator for thermal ranking of low-cost wall materials in hot-dry climates using finite difference methods.

MyMentorLLM: multimodal LLM simulation environment for psychotherapy training with voice/text patients, trainees, and expert supervisors.

Contextual Deconvolution method for demand forecasting in retail using kernel-modulated operators to separate promotional shocks from baseline trends.

Controlled benchmark analyzing how LoRA adaptation site affects learning geometry across lexical, factual, causal, and procedural objectives in Transformers.

CoRT: token-level credit weighting for rubric-guided policy optimization in language models, enabling fine-grained attribution across generated spans.

OrchBench: simulation-based benchmark for evaluating multi-agent orchestration plans in isolation without conflating with worker/tool/environmental noise.

Analysis of outcome skew in Stockfish-evaluated equal chess positions vs. human play on Lichess, showing position-specific bias independent of ratings.

Policy analysis: general-purpose AI governance frameworks for public services risk failure due to GPAI properties (generality, accessibility, low cost) undermining safety preconditions.

KQFuzz: LLM-based knowledge-guided fuzzer for quantum libraries using codebase context to improve test generation efficiency and flexibility.

Perplexity has expanded its agentic Personal Computer tool to Windows, allowing computers running the world's most popular OS to be used as a locally run AI system. Like the Mac version that Perplexity launched in April, Personal Computer for Windows operates like a "general-purpose digital worker" that can access local files and apps to perform actions on your behalf, such as creating documents and updating spreadsheets. This launch builds on Personal Computer integrations that Perplexity launched for Microsoft's 365 workspace apps and Teams virtual meeting software in May. Personal Computer...

Survey of instruction-based image editing spanning task taxonomy, training data methods, and architectural evolution from GAN to diffusion/autoregressive models.

OmniPhys benchmark (1,551 samples) diagnoses physical commonsense violations in text-to-image models via knowledge graphs and addresses gradient hallucination in prompt optimization.

F(AI)2R extends AI provenance tracking to executable agent skills via aiprov (PROV-O extension), enabling auditable AI contributions in research artifacts.

Study uses spectral dataset properties (entropy, rank, autocorrelation) as empirical priors to guide 1D-CNN architecture design for NIR chemometrics across 25 regression tasks.

AIriskEval-edu audits pedagogical risks in AI-generated explanations across five dimensions (factual accuracy, depth, bias, relevance, student-appropriateness) with evidence spans.

Construction-Driven Injection embeds linguistically-grounded code-mixing fingerprints in LLMs to create verifiable ownership signals resistant to accidental activation and perplexity filtering.

Human-in-the-loop corpus simplifies scientific paper summaries via GPT-4o-mini with iterative refinement by non-specialist STEM readers for cross-disciplinary accessibility.