Loading…
Browse by type
Benchmark reads
Benchmark posts on Yomu, summarised newest first.
- 172 posts
- 20 companies
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
Benchmark posts
Better AI models enable more ambitious work
An eight-month study of developers at 500 companies using Cursor examined whether improved AI mainly increases existing work or enables tasks previously out of reach. The research, conducted with Professor Suproteem Sarkar of the University of Chicago Booth School of Business, covered releases of Opus 4.5 and GPT-5.2, finding average weekly messages per user rose 44%. Developers initially used better models for more work of similar complexity, then after a 4–6-week lag shifted toward harder tasks: low-complexity messages grew 22%, while high-complexity messages grew 68%, mostly in the final six weeks. Growth varied by sector, led by media and advertising at 54%, software and developer tools at 47%, and finance and fintech at 45%. Documentation, architecture, code review, and learning grew more than UI and styling, suggesting expanding AI-generated codebases increase demand for management, understanding, and review while opening broader productive opportunities.
New Research Reveals Overcoming Legacy Tech Issues Key to AI Success
An IDC survey of 1,400 organizations across eight Asia/Pacific markets reveals that legacy architectures and accumulated data debt severely hinder enterprise artificial intelligence initiatives. While 43% of surveyed organizations cite legacy systems as a major obstacle to application development, 95% report project delays and 90% have experienced failed modernization attempts. In contrast, high-performing organizations treat modernization as a continuous discipline across people, processes, and cloud-centric data platforms capable of processing both structured and unstructured workloads. These leading enterprises achieve nearly three times more digital revenue growth than peers still constrained by technical debt. IDC notes that addressing technical and data debt proactively avoids escalating AI failure rates and delivers superior return on investment compared to superficial lift-and-shift migrations.
Agentic reasoning in practice: Making sense of structured and unstructured data
Enterprise questions that connect structured records with unstructured text are difficult for single-turn retrieval systems to answer reliably. The post evaluates Databricks Agent Bricks Supervisor Agent (SA), a declarative agent builder built on the internal aroll framework, across STaRK’s Amazon, MAG, and Prime domains and six KARLBench tasks. SA decomposes questions, routes sub-questions to Genie or Knowledge Assistant, combines results, and can reformulate queries or use SQL JOINs when initial retrieval is incomplete. It outperforms rerun SoTA baselines by 4% on STaRK-Amazon, 21% on MAG, and 38% on Prime, while delivering a 23% improvement on FinanceBench and consistent gains across KARLBench; the authors attribute the results to multi-step tool selection, reflection, and precise instructions rather than custom code.
Speeding up GPU kernels by 38% with a multi-agent system
The post reports a three-week experiment in which a multi-agent system optimized 235 CUDA kernel problems generated from workloads spanning more than 124 production open-source models. Using SOL-ExecBench on 27 NVIDIA Blackwell 200 GPUs, the system used a planner to distribute and rebalance work, while autonomous workers benchmarked, debugged, and refined kernels without developer intervention. It ran separate CUDA C with inline PTX and CuTe DSL runs, reaching assembly-level optimization and learning a documented higher-level API. Against a single-agent optimized PyTorch baseline, it outperformed 149 of 235 problems, achieving a 1.38x geometric mean ratio, while 45 solutions exceeded 2x. The highlighted cases included an 84% geomean gain over a human FlashInfer baseline for paged-prefill attention, a 39% gain for NVFP4 MoE linear, and a GEMM reaching 86% of a tuned cuBLAS baseline; median SOL was 0.56, leaving room for improvement.
Better MoE model inference with warp decode
MoE inference commonly organizes decode around experts, but the post describes “warp decode,” an output-centric kernel for small-batch generation on NVIDIA Blackwell GPUs. Instead of staging tokens into expert-major buffers, each warp owns one output scalar, streams the required weight rows, accumulates contributions from all eight routed experts in FP32, and uses warp-level reductions; the gate/up and down projections are compressed into two kernels. This removes padding, scatter, combine, activation gather, and per-expert output buffers, while keeping BF16 activations and avoiding intermediate quantization. On B200s, the approach delivered 1.84x higher decode throughput, sustained 3.95 TB/s at batch size 32, and produced outputs 1.4x closer to a full FP32 reference.
PostgresBench: A Reproducible Benchmark for Postgres Services
PostgresBench is a public, reproducible benchmarking suite created to compare transactional performance across managed PostgreSQL providers. The benchmark uses the standard pgbench tool with a TPC-B-like write-intensive workload executing across 256 clients and 16 threads for ten-minute durations. Evaluators tested five services—including Postgres managed by ClickHouse, AWS Aurora, AWS RDS, Neon, and Crunchy Bridge—under 100 GB and 500 GB scale factors using default configurations in the us-east-2 region. Across tested configurations, Postgres managed by ClickHouse delivered higher transactions per second and lower P99 latencies due to physically co-located NVMe primary storage. These results demonstrate that remote network storage layers, such as EBS volumes, introduce significant latency overhead during sustained write and fsync operations.
Training mRNA Language Models Across 25 Species for $165
OpenMed describes an end-to-end protein engineering pipeline that connects ESMFold structure prediction, ProteinMPNN sequence design, and custom mRNA codon optimization. The codon-language-model study compares CodonBERT, ModernBERT, and several CodonRoBERTa variants using 250,000 E. coli coding sequences, a 69-token codon vocabulary, masked-language modeling, and training on four 80GB A100 GPUs. CodonRoBERTa-large-v2 achieved perplexity 4.10, CAI Spearman correlation 0.404, and 7.7% synonymous recovery, while ModernBERT-base reached perplexity 26.24 and correlation 0.070. The work then trained a multi-species suite on 381,283 sequences spanning 25 species, producing four production models in 55 GPU-hours and documenting planned extensions such as larger training data and additional prediction heads.
Boost Training Goodput: How Continuous Checkpointing Optimizes Reliability in Orbax and MaxText
Orbax and MaxText introduce continuous checkpointing to balance training reliability and performance when hardware failures or preemption events interrupt long-running jobs. Fixed checkpoint intervals create a trade-off: infrequent saves risk substantial lost work, while frequent saves can block or bottleneck training, particularly over unstable networks. MaxText enables asynchronous checkpointing that attempts a new save only after the previous background request is fulfilled; its continuous mode overrides checkpoint_period, and max_num_checkpoints_to_keep can limit storage use. A llama-3.1-70B benchmark on two v5p-128 slices found smaller P50 checkpoint intervals, but average training step time increased because of more frequent device-to-host transfers; the analysis estimated resource conservation against mean-time-between-failure assumptions. Orbax also supports minimum intervals and customizable preservation policies, while co-locating storage with training is recommended because cross-metro bandwidth can reduce checkpointing speed and reliability.
A New Framework for Evaluating Voice Agents (EVA)
EVA is an end-to-end framework for evaluating conversational voice agents across task accuracy and spoken conversational experience. It simulates multi-turn live-audio conversations with a user simulator, the evaluated voice agent, deterministic tool execution, validation, and metrics derived from recordings, transcripts, and tool-call logs. The framework reports EVA-A for Accuracy and EVA-X for Experience, combines code-based and LLM- or LALM-as-Judge metrics, and uses pass@3 and pass^3 across three trials to capture peak performance and consistency. Its initial release contains 50 reproducible English-language airline scenarios and 15 tools covering rebooking, itinerary changes, cancellations, standby, and vouchers. Results for 20 cascade and audio-native systems show a consistent tradeoff: agents stronger at task completion tend to provide worse experiences, while a substantial gap between pass@3 and pass^3 indicates inconsistent success.
How we compare model quality in Cursor
Cursor describes a hybrid online-offline process for comparing coding-agent quality as developer tasks become longer, more complex, and multi-file. Its offline CursorBench suite uses tasks derived from real Cursor sessions, paired with ground-truth solutions through Cursor Blame, and evaluates correctness alongside code quality, efficiency, and interaction behavior; the post focuses on correctness. Tasks are intentionally short and underspecified, span larger codebases and multi-workspace environments, and are scored with agentic graders, while correctness is plotted against median completion tokens to show the compute/latency trade-off. Controlled live-traffic experiments supplement the suite, catching regressions that offline grading misses, and the post reports that CursorBench separates frontier models more clearly and its rankings track developer experience more closely than public benchmarks.
Keep the Tokens Flowing: Lessons from 16 Open-Source RL Libraries
At-scale reinforcement-learning training is constrained by long, variable data-generation workloads that leave training GPUs idle, especially for reasoning and agentic models. The surveyed architecture separates inference and training onto distinct GPU pools, connects them through a rollout buffer, and synchronizes weights asynchronously so both stages can proceed concurrently. Sixteen open-source libraries are compared across orchestration, buffering, weight synchronization, staleness, partial rollouts, LoRA, and distributed backends; Ray appears in 8 of 16 orchestration implementations, while NCCL broadcast is the default weight-transfer method. The proposed TRL async trainer favors a lightweight bounded queue with per-token model_version metadata, packed NCCL transfers, and experimental support for prefix-resume or abort-and-retry handling of rollouts interrupted by weight updates.
Financial Benchmarks
Ramp describes how it benchmarks large language models (LLMs) used in financial products against day-to-day tasks such as invoice extraction, financial-statement OCR, policy review, accounting autocoding, compliance detection, and fund routing. These benchmarks combine task-specific metrics with cost, latency, reasoning effort, human decisions, historical context, and ground-truth datasets where available. For contextual invoice OCR, perfect extraction requires every field to match the user's final bill, while financial-statement OCR uses a 1% relative-error threshold against transcriptions from over 500 real P&L documents. Results show application-specific trade-offs: Gemini 3 Flash is described as a cost-efficient leader for several visual and financial tasks, while Claude models lead some high-accuracy or low-miss-rate settings, and behavior can differ within one provider. The framework emphasizes Pareto trade-offs and continuous testing, with plans to expand from benchmarking current performance to hill climbing on capabilities.
PRX Part 3 — Training a Text-to-Image Model in 24h!
PRX Part 3 reports a 24-hour speedrun for training a text-to-image diffusion model, combining techniques that performed well in earlier experiments under a budget of about $1,500 on 32 H200 GPUs. The recipe uses x-prediction in pixel space with 32-pixel patches and a 256-dimensional initial bottleneck, followed by LPIPS and DINOv2 perceptual losses, TREAD token routing, REPA alignment with DINOv3, and Muon for matrix parameters while Adam handles the rest. Training ran for 100,000 steps at 512px with batch size 1,024, then 20,000 steps at 1024px with batch size 512 and no REPA, across three synthetic datasets. The resulting model is described as usable, with strong prompt following and consistent aesthetics, though texture glitches, anatomy issues, and difficulty with hard prompts remain. The post attributes these limitations mainly to undertraining and limited data diversity and presents the recipe as a foundation for further scaling.
2,000 robots walk into a shop: Simulated A/B testing (2026) - Shopify
Shopify’s SimGym addresses the difficulty of evaluating storefront changes when merchants lack enough traffic for conventional A/B tests, by simulating shopping sessions with LLM-guided robots. Each robot runs in a cloud Chromium browser with a persona, budget, and intent, while a counterpart tests the alternate theme; the system uses real page representations rather than a mocked DOM. To support bursty, sequential workloads with 89K–127K accumulated tokens per session, Shopify self-hosted gpt-oss-120b on 48 NVIDIA B200 GPUs, combining MXFP4 expert quantization, FP8 KV-cache, FlashInfer kernels, async scheduling, and guided decoding. Production measurements report 80K to 88K tokens/second per B200 from async scheduling, a 57% reduction in HTTP/gRPC overhead, and MIG experiments yielding about 20% lower latency and 12% higher daily throughput without quality regression; speculative decoding is validated but queued.
IBM and UC Berkeley Diagnose Why Enterprise Agents Fail Using IT-Bench and MAST
IBM Research and UC Berkeley analyze how agentic LLM systems fail in enterprise IT automation, including incident triage, logs and metrics queries, and Kubernetes actions across long-horizon tool loops. To move beyond benchmark success rates, the analysis applies MAST (Multi-Agent System Failure Taxonomy) to execution traces, converting failures into structured signatures across system design, inter-agent misalignment, and task verification. It covers 310 ITBench SRE traces from Gemini-3-Flash, Kimi-K2, and GPT-OSS-120B, whose mean recalls were 75.5%, 28.6%, and 12.4%, respectively. Gemini averaged 2.6 failure modes per failed trace, compared with 4.7 for Kimi and 5.3 for GPT-OSS-120B, while FM-3.3, Incorrect Verification, was identified as the strongest predictor across models; the analysis recommends external verification, explicit termination controls, and clarification branches.
Training Design for Text-to-Image Models: Lessons from Ablations
The second part of a series on training efficient text-to-image models from scratch documents training-design ablations for the PRX model. The authors establish a Flow Matching baseline (PRX-1.2B, 100K steps, 256×256, AdamW, GemmaT5, RoPE, 1M MidJourneyV6 synthetic images) and benchmark interventions against FID, CMMD, DINO-MMD, and throughput across four buckets: Representation Alignment, Training Objectives, Token Routing and Sparsification, and Data. REPA representation alignment boosts early convergence when used as a burn-in, while the Muon optimizer lowered FID from 18.20 to 15.55 and improved CMMD and DINO-MMD. Storing denoiser weights in BF16 instead of using BF16 autocast raised FID to 21.87, and token routing methods such as TREAD/SPRINT were minor at 256² but became major wins at higher resolutions. Long captions proved critical, synthetic versus real data shifted texture versus structure, and upcoming work releases the full training code alongside a 24-hour end-to-end speedrun.
The AI deployment gap is widening, don’t get left behind
The 2026 Customer Service Transformation Report examines how organizations are adopting AI in customer service, based on a survey of over 2,400 global customer service professionals. It identifies a widening difference between narrow deployments that handle simple questions or workflow steps and mature deployments that integrate AI into critical operations, assign it more responsibility, and improve it continuously. Although 62% of respondents report improved customer-service metrics after implementing AI, mature teams report this at 87%, while only 10% say deployment is mature. The report also finds mature deployment associated with higher quality and consistency, clearer ROI measurement, and less time spent handling support volume; 43% report higher quality and consistency. With customer experience becoming the top priority for 58% in 2026, support is positioned as a proving ground for expanding AI into customer success, marketing, and sales, with 52% planning that scale.
Alyah ⭐️: Toward Robust Evaluation of Emirati Dialect Capabilities in Arabic LLMs
Standard Arabic language model evaluations predominantly assess Modern Standard Arabic while overlooking regional spoken dialects that feature distinct syntax, vocabulary, and cultural grounding. To address this evaluation gap, researchers created Alyah, a benchmark comprising 1,173 manually curated multiple-choice questions collected from native Emirati speakers across seven cultural and linguistic categories. Distractor answer choices were synthetically generated using language models and manually reviewed for semantic plausibility. Evaluating base and instruction-tuned Arabic-native, adapted, and multilingual models revealed that instruction-tuned systems consistently outperformed base variants, with falcon-h1-arabic-7b-instruct achieving the top instruction accuracy at 82.18%. However, questions involving implicit meanings, rare expressions, and specific dialect phenomena remained difficult across all tested model families and sizes.
AssetOpsBench: Bridging the Gap Between AI Agent Benchmarks and Industrial Reality
AssetOpsBench is an evaluation framework designed to benchmark agentic AI systems within industrial Asset Lifecycle Management operations, such as chillers and air handling units. The benchmark incorporates 2.3 million sensor telemetry points, 4,200 work orders, 53 structured failure modes, and over 140 curated scenarios spanning four agent roles. Systems are assessed across six qualitative criteria, including task completion, retrieval accuracy, result verification, sequence correctness, clarity and justification, and hallucination rate. Evaluation traces are analyzed using a trajectory-level diagnostic pipeline called TrajFM, which extracts and clusters failure patterns. In community tests covering over 300 agents, no evaluated model family reached the 85-point readiness threshold.
Open ASR Leaderboard: Trends and Insights with New Multilingual & Long-Form Tracks
The Open ASR Leaderboard expanded its evaluation framework by introducing dedicated multilingual and long-form transcription tracks across eleven datasets. Benchmarking more than sixty open- and closed-source models from eighteen organizations revealed clear operational trade-offs between accuracy and inference speed. Systems combining Conformer encoders with large language model decoders achieved the lowest word error rates in short-form English speech recognition. In contrast, connectionist temporal classification and token-and-duration transducer decoders yielded tenfold to hundredfold improvements in throughput with only slight accuracy penalties. While closed-source commercial systems maintained performance leads in long-form audio transcription, specialized single-language models consistently outperformed generalized systems supporting hundreds of languages.
Top topics
- Architecture834
- Performance816
- LLMs660
- Machine Learning622
- Open Source563
- AI Agents553
- AI470
- Developer Experience448
Weekly digest
The best engineering reads, once a week.