ADIAS: Automated Design of Interactive Agentic Systems
ADIAS targets a subtle but costly flaw in agent-building pipelines: most methods organize progress around candidate agents, so “repair progress” is rebuilt implicitly each round. The new issue-centric approach carries forward a persistent issue state so optimization can focus on stable targets, not re-learn what changed. Across five interactive benchmarks, ADIAS beats the strongest baseline by 25.2% on average and keeps gains across four backbone models.
Build Low-Latency Multilingual Voice Agents: Open Weights & Full Deployment Control with NVIDIA Magpie TTS
NVIDIA’s Magpie TTS is presented as a practical building block for multilingual voice agents where latency and deployment control matter more than polished demos. The emphasis is on open-weight availability and the ability to run full stacks you can tune, rather than treating speech as a black box. For agent builders, that means faster iteration cycles and more predictable production behavior.
What building an AI-native finance function taught me
OpenAI’s finance-focused write-up frames AI adoption as a chain from forecasting and automation to governance, with “controls” and traceability treated as first-class requirements. The piece argues that AI ROI comes not just from model performance, but from embedding decision support into workflows that are auditable. It’s a blueprint for moving from experimental copilots to operational finance systems.
What happens to Bose when headphones become AI?
In a conversation with Bose CEO Lila Snyder, the discussion pivots from headphones as a product category to audio technology as an ingredient used across devices. Snyder outlines how Bose is expanding into B2B licensing and what that means for preserving sound quality as AI wearables and conversational interfaces become mainstream. The interview also connects on-device constraints like power and compute to why “tiny AI” still needs to deliver real audio clarity.
Learning to Predict Middle-Layer Attention in MLLMs for Visual Token Pruning
This work attacks the expensive part of multimodal inference: too many visual tokens. Instead of assuming a fixed “best” attention layer for pruning, it predicts the right middle-layer attention per sample using a question-contrastive teacher selection scheme and then distills it into a lightweight predictor. Across 10 benchmarks, it keeps 97.5% of unpruned performance while pruning down to 5.56% of visual tokens, claiming a 3.09x end-to-end speedup.
Model ML completes finance work more efficiently with GPT-5.6 Sol
OpenAI highlights Model ML’s workflow improvements after adopting GPT-5.6 Sol for turning research and analysis into editable, traceable outputs like PowerPoint decks and Excel workbooks. The thrust is practical: less manual transcription and formatting, while keeping artifacts reviewable and aligned to the underlying work. For teams drowning in spreadsheet drift and deck rewrites, this positions “agentic” finance labor as a measurable productivity lever.
WebGrader: Training LLMs for Web Development with Self-Evolving Programmatic Grader
WebGrader takes aim at a core bottleneck in web agent RL: reward design that’s either too expensive to write by hand or too unreliable by guessing before the decisive browser state is reached. It auto-derives interaction flows, executes them as “Flow Contracts” in a live browser, and uses the observed transition as the reward signal. On WebGen-Bench, it trains an 8B policy to 52.01% functional success and reports gains over other reward strategies and strong competitive baselines.
Efficient Knowledge Distillation (efficient-knowledge-distillation)
This post frames knowledge distillation as a deployment problem: how do you compress capabilities without breaking the behaviors you care about at runtime. The emphasis is on making distillation cheap enough to run at scale, so teams can repeatedly refresh smaller models as data and requirements change. It’s pitched for practitioners who need performance-per-dollar, not just academic “student matches teacher” results.
These startups are chasing the next big thing in LLMs
MIT Technology Review surveys how new companies are aiming beyond the baseline “bigger model” race, homing in on infrastructure and agentic capabilities that make LLMs useful in real environments. The reporting highlights why distribution, orchestration, and evaluation are becoming as important as raw modeling. If you’re watching the LLM startup landscape, this provides a map of where founders think the next scalable wedge will be.
Ford’s new AI assistant can check your fuel levels and tire pressure
Ford is rolling out an AI assistant inside its Ford and Lincoln apps that can answer vehicle-specific questions, from fuel calculations to towing and service needs. Because it’s tied to the customer’s actual vehicle data, the assistant can respond with details like fuel levels, cargo capacity, and towing capability. The follow-on plan includes voice-powered interactions, moving this from “manual lookup” into conversational upkeep.