1. Measuring the Environmental Impact of Delivering AI at Google Scale
Google presents firstâparty, production measurements of AI servingâs environmental impact for Gemini Apps. Using a fullâstack boundary that includes accelerator power, host CPU/DRAM, provisioned idle capacity, and dataâcenter overhead, the team finds the median Gemini text prompt is far lower impact than many public estimates and shows rapid efficiency gains over one year.
What was actually measured â A comprehensive âserving AI computerâ boundary: active AI accelerators, host CPU/DRAM, idle machines kept for reliability/latency, and dataâcenter overhead via PUE. Networking, endâuser devices, and training are excluded. Figure 1 illustrates the boundary choices.
Key numbers for a median text prompt (May 2025) â 0.24 Wh energy, 0.03 gCO2e marketâbased emissions, 0.26 mL water. This is roughly less energy than watching TV for 9 seconds and about five drops of water. Table 1 breaks down contributions: accelerators 0.14 Wh, host 0.06 Wh, idle 0.02 Wh, overhead 0.02 Wh.
Why do many estimates differ? Narrow acceleratorâonly approaches undercount. The paper shows a 1.72Ă uplift from accelerator energy to total serving energy when you include host, idle, and overhead. In a benchmarkâlike âexisting approach,â the same prompt would appear as 0.10 Wh.
Yearâoverâyear efficiency gains â From May 2024 to May 2025, median perâprompt emissions fell 44Ă driven by software/model improvements (33Ă energy reduction, including 23Ă from model changes and 1.4Ă from utilization), cleaner electricity (1.4Ă), and lower amortized embodied emissions (36Ă).
How to use these metrics â The authors argue for median perâprompt reporting to avoid skew from long or lowâutilization prompts, and for standardized, fullâstack boundaries so providers can compare models and target the biggest levers across software, hardware, fleet utilization, siting, and cleanâenergy procurement.
2. Avengers-Pro: Beyond GPT-5
A lightweight testâtime router that embeds each query, maps it to semantic clusters, and selects one LLM from an ensemble to optimize accuracy vs cost. On six hard benchmarks and eight leading models, it beats the best single model at a similar cost and traces a Pareto frontier of accuracy for any budget.
What it is: Embed â cluster â score. Queries are embedded, assigned to nearest clusters, then routed to the model with the highest performanceâefficiency score controlled by a weight α. Only one model answers each query.
Why it matters: With comparable cost, AvengersâPro outperforms GPTâ5âmedium by about 7% average accuracy; with comparable accuracy, it reduces cost by about 27%. Hitting ~90% of GPTâ5âmediumâs accuracy costs ~63% less.
Setup: Ensemble of 8 models (GPTâ5âchat/medium, Claudeâ4.1âopus/Sonnetâ4, Geminiâ2.5âpro/flash, Qwen3 235B and thinking). Evaluated on GPQAâDiamond, HLE, ARCâAGI, SimpleQA, LiveCodeBench, and ÏÂČâbench. Pricing via OpenRouter informs perâcluster cost scoring.
Knobs that matter: α tunes performance vs efficiency; accuracy rises fast until ~0.6 while cost stays low until ~0.4, then climbs. Implementation uses kâmeans with k=60, Qwen3âembeddingâ8B (4096âd) and topâp=4 nearest clusters at inference.
Routing behavior: At low α, the router favors cheaper Qwen3 variants; as α grows, it shifts to GPTâ5âmedium and other stronger but costlier models.
3. Chain-of-Agents
OPPO proposes training single models to natively behave like multiâagent systems, coordinating roleâplaying and tool agents endâtoâend. They distill strong multiâagent frameworks into CoA trajectories, then optimize with agentic RL on verifiable tasks. The result: AFMs that solve complex web, code, and math problems with less overhead and new stateâofâtheâart results.
Paradigm shift: CoA generalizes ReAct/TIR by dynamically activating multiple roles and tools within one model, preserving a single coherent state while cutting interâagent chatter.
Training recipe: 1) Multiâagent distillation turns successful OAgents runs into CoAâformatted traces with planning, tool calls, observations, and reflection, filtered for difficulty and quality; 2) Agentic RL targets hard queries where tools matter, with simple binary rewards via LLMâasâJudge for web tasks and executable or exactâmatch rewards for code/math.
Main results: With Qwenâ2.5â32B backbones, AFM sets new pass@1 on GAIA 55.3, BrowseComp 11.1, HLE 18.0, and leads WebWalker 63.0; it also tops multiâhop QA suites across sizes.
Code + math: AFMâRLâ32B reaches AIME25 59.8, MATH500 94.6, OlympiadBench 72.1, and LiveCodeBench v5 47.9, beating prior TIR methods including ReTool and Reveal.
Efficiency and robustness: Compared to traditional multiâagent systems, AFM cuts inference tokens and tool calls substantially; the paper reports an 84.6% token cost reduction while staying competitive. It also generalizes to unseen tools better when strict formatting is required.
Testâtime scaling: Bestâofâ3 and pass@3 markedly boost AFM, e.g., GAIA 69.9 and HLE 33.2, closing the gap with larger proprietary agent stacks.
Open source: Models, data, and training code are released to spur research on agent models and agentic RL.
4. Has GPT-5 Achieved Spatial Intelligence?
This report introduces a unified view of spatial intelligence (SI) for multimodal models and evaluates GPTâ5 and strong baselines across eight fresh SI benchmarks. GPTâ5 leads overall but is still short of human skill, especially on mentally reconstructing shapes, changing viewpoints, and deformation/assembly tasks.
Unified SI schema and fair eval setup. The authors consolidate prior work into six core SI capabilities (Metric Measurement, Mental Reconstruction, Spatial Relations, Perspectiveâtaking, Deformation & Assembly, Comprehensive Reasoning) and standardize prompts, answer extraction, and metrics to reduce evaluation variance across datasets.
Broad benchmark sweep, heavy compute. Eight recent benchmarks (e.g., VSIâBench, SITE, MMSI, OmniSpatial, MindCube, STARE, CoreCognition, SpatialViz) are used with unified protocols; results reflect >1B tokens of evaluation traffic.
GPTâ5 sets SOTA but not humanâlevel SI. GPTâ5 tops aggregate scores and sometimes reaches human parity on Metric Measurement and Spatial Relations, yet shows significant gaps on Mental Reconstruction, Perspectiveâtaking, Deformation & Assembly, and multiâstage Comprehensive Reasoning.
Hard SI narrows the closed vs open gap. While proprietary models win on average, their advantage evaporates on the hardest SI categories; several openâsource systems perform similarly, far from human ability on MR/PT/DA/CR. NonâSI portions (e.g., CoreCognitionâs Formal Operation) can reach nearâhuman levels.
Qualitative analysis exposes failure modes. Case studies show prompt sensitivity for novelâview generation, blind spots with perspective effects and size constancy, persistent failures on paperâfolding/assembly, and difficulty inferring occluded objects during counting.
5. ComputerRL
A framework for autonomous desktop agents that unifies API calls with GUI actions, plus a scalable RL stack and a training recipe (Entropulse) that alternates RL and SFT to sustain exploration. Evaluated on OSWorld, it sets a new SOTA with strong gains in efficiency and robustness.
APIâGUI action space. Moves beyond humanâcentric GUIs by combining programmatic APIs with direct GUI control. LLMs help autoâgenerate appâspecific APIs via requirement analysis, implementation, and unit tests, lowering the cost of adding new tools.
Massively parallel desktop env. A refactored Ubuntu VM cluster (qemuâinâdocker + gRPC) delivers thousands of concurrent instances with improved stability, monitoring, and AgentBenchâcompatible interfaces, enabling largeâscale online RL.
Fully asynchronous RL. Built on AgentRL with decoupled actors/trainers, dynamic batching, and bounded replay to reduce offâpolicy bias and maximize GPU utilization during longâhorizon desktop rollouts.
Entropulse training. After an initial stepâlevel GRPO phase with verifiable, ruleâbased rewards, successful rollouts are distilled via SFT to restore entropy, then RL resumes, yielding higher rewards and sustained improvements.
Results and analysis. AUTOGLMâOSâ9B reaches 48.1% on OSWorld and 47.3% on OSWorldâVerified, outperforming OpenAI CUA o3, UIâTARSâ1.5, and Claude Sonnet 4, while using up to oneâthird the action steps of strong baselines; ablations show APIâGUI and multiâstage training drive the gains. Error sources cluster into multiâapp coordination, vision, and operational slips.
6. Full-Stack Fine-Tuning for the Q Programming Language
Presents an open-source blueprint for adapting large language models to niche programming domains, with Q (used in quantitative finance) as the test case. The team builds a benchmark, curates data, and trains Qwen-2.5 models with pretraining, supervised fine-tuning, and reinforcement learning. Their largest model surpasses Claude Opus-4 by nearly 30% on Q-LeetCode tasks, and even the smallest model beats GPT-4.1.
Key points:
Benchmark creation â Introduced the first LeetCode-style dataset for Q, comprising 678 problems with automated and human-verified solutions.
Model performance â Trained models from 1.5B to 32B parameters; all exceed GPT-4.1, and the 32B reasoning variant achieves 59% pass@1, +29.5% over Claude Opus-4.
Training pipeline â Multi-stage process: domain-adaptive pretraining on Q code, supervised fine-tuning on curated tasks, and reinforcement learning with programmatic rewards.
Lessons learned â Robust evaluation harness and data quality are critical; reward hacking is pervasive without careful separation of solution/test generation; large models (â„14B) are essential for meaningful gains.
Limitations â The dataset uses âpythonicâ Q (algorithmic tasks), not typical finance workloads (queries, time-series analytics), so real-world performance may differ.
7. As Generative Models Improve, People Adapt Their Prompts
A large online experiment (N = 1,893) compares DALL·E 2, DALL·E 3, and DALL·E 3 with automatic prompt revision on a 10âattempt image replication task. DALL·E 3 improves outcomes not only because the model is better, but because people change how they prompt when the model is stronger.
Headline effect: Relative to DALL·E 2, DALL·E 3 yields images closer to targets by âCoSim = 0.0164, about z = 0.19 SD, with the gap widening across attempts.
Behavior adapts to capability: Without knowing which model they used, DALL·E 3 participants wrote longer prompts (+24%, +6.9 words on average) that added descriptive content, and their prompts became more semantically similar to each other over iterations.
Decomposed gains: About half of the improvement is due to the model itself and about half to usersâ adapted prompting. The ATE splits into a model effect of âCoSim â 0.00841 (51%) and a prompting effect of âCoSim â 0.00788 (48%).
Prompt revision caveat: Automatic LLM prompt revision helps over DALL·E 2, but it cuts the DALL·E 3 advantage by ~58% and can misalign with user goals.
Takeaway: As models advance, users naturally supply richer, more consistent prompts that the newer models can realize more effectively. Prompting remains central to unlocking capability gains.
8. Retrieval-Augmented Reasoning with Lean Language Models
A domain-tuned pipeline that fuses RAG and reasoning into a single small-footprint model. The team distills reasoning traces from a frontier model into Qwen2.5 variants, uses summarization to keep context small, and shows that a 32B local model approaches frontier accuracy on an NHS AâtoâZ clinical QA task.
Method in one picture. The system marries a dense retriever (SentenceâTransformers + Chroma/FAISS) with Qwen2.5âInstruct models; retrieval can be invoked as a tool inside a conversational agent.
Lean + private by design. Built to run in secure or airâgapped settings using open models; integrates reasoning with retrieval to reduce hallucinations while keeping data onâprem.
Data and compression. On ~1k NHS condition pages, the team generates synthetic queries, retrieves full documents, then summarizes them to shrink input by 85% (avg trace length from ~74,641 to ~7,544 tokens) before fineâtuning.
Retriever wins, then reasoner adds. Summaries beat full pages for retrieval (p@5: 0.76 vs 0.68). With k=5 retrieved docs, condition accuracy caps at 0.76 upstream; within that cap, Qwen2.5â32B jumps from 0.38 to 0.54 with RAG, and to 0.56 after reasoning distillation. Frontier baselines with RAG land around 0.56â0.57.
Small models, big gains. Distilled ât0â models from 1.5Bâ32B retain strong condition accuracy with k=5 (e.g., 1.5B at 0.53; 32B at 0.56), narrowing the gap to frontier models while fitting in 3â64âŻGB GPU memory. The study highlights that reasoning distillation especially lifts the smallest models.
Practicality. Training reused s1âstyle SFT with long context on accessible hardware (e.g., 16ĂA100 80âŻGB; ~80 GPUâhours for the 32B run), and ships a simple Svelte frontend with hidden âreasoning traceâ toggles for auditability.
9. Parallel Text Generation
This survey reviews parallel text generation methods that overcome the sequential limits of autoregressive decoding by enabling faster inference. It categorizes AR-based and non-AR-based approaches, analyzes trade-offs in speed, quality, and efficiency, and highlights recent advances, open challenges, and future research directions.
10. Open Foundations for Compute-Use Agents
This paper introduces an open-source framework for computer-use agents, featuring a large-scale dataset (AgentNet), annotation infrastructure, and reflective reasoning pipelines. Its 32B model sets a new SOTA on OSWorld-Verified with 34.8% success, surpassing OpenAIâs GPT-4o, and all resources are released to advance open CUA research.