DeepSeek made its V4 Pro price discount permanent, the day's clearest market signal — sustained price pressure from the open-weight frontier keeps compressing what proprietary labs can charge. The research feeds pushed agents toward the surfaces businesses actually use: Spreadsheet-RL trains agents on real Excel and Sheets tasks, TerminalWorld reverse-engineers evaluation from in-the-wild terminal work, and π-Bench measures proactive personal-assistant agents over long horizons. Clinical agents got more rigorous with multimodal evidence-seeking and clinical-event prediction. Efficient attention and KV serving stayed busy — full-to-sparse attention transfer in a hundred training steps, a second-generation gated DeltaNet, and service-aware cache compression for disaggregated serving.
DeepSeek made its V4 Pro price discount permanent. Open-weight frontier pricing keeps dropping, and each permanent cut tightens the band proprietary labs can charge for comparable capability — a direct tailwind for cost-sensitive SMB builders.
DeepSeek converted its temporary V4 Pro price discount into permanent pricing, locking in another cut to frontier-grade API costs (534 HN points).
Three benchmarks pull agent evaluation onto the surfaces businesses actually run on: spreadsheets, terminals, and proactive long-horizon assistance. The shared move is reverse-engineering tasks from real-world usage rather than synthetic sandboxes — the same controlled-to-realistic shift seen all week.
Trains agents with RL on realistic Excel and Google Sheets tasks — the data-work surface most businesses actually live in.
A data engine that auto-reverse-engineers high-fidelity evaluation tasks from in-the-wild terminal usage.
Benchmark for proactive personal-assistant agents across long-horizon, everyday workflows.
Uses RL to orchestrate hierarchical ensembles of models and skills for autonomous agents.
Healthcare AI research moved past assuming evidence is handed to the model: ClinSeekAgent automates multimodal evidence-seeking for clinical reasoning, and a separate effort trains LLMs to predict clinical events from longitudinal notes. Both respond to the deployment-risk story the field has been living this month.
Agentic clinical reasoning that actively seeks multimodal evidence rather than assuming it is already in context.
Converts longitudinal clinical notes into supervision for predicting how patients evolve over time.
Efficiency work continued across the stack: converting full attention to sparse in a hundred training steps, a second-generation gated DeltaNet that decouples erase and write in linear attention, and service-aware KV-cache compression for disaggregated serving.
Converts a full-attention model to efficient sparse attention in roughly a hundred training steps, keeping quality.
Separates erase and write operations in linear attention's recurrent state for better sequence modeling at linear cost.
Service-aware KV-cache compression that cuts communication cost in disaggregated (prefill/decode-separated) serving.
Two papers continue the week's RLVR refinement thread — discriminative token-level credit assignment and unsupervised process reward models — both aimed at giving training a signal more precise than one reward per rollout.
Assigns credit to individual tokens within an RLVR rollout rather than rewarding all tokens equally.
Builds step-level process reward models without step-level human labels.