groundy
articlessearch

models & research

128 articles·rss

Where architecture, training tricks, and eval methodology meet the marketing layer — separating durable progress in foundation models from leaderboard theater that quietly falls apart under load.

latest
A copper speaking horn projects toward a glass-covered precision gauge whose internal mechanism and needle disagree, with calibration weights below.

Can You Trust LLM Confidence Scores? Verbalized vs Logprob Signals

A preprint reports that verbalized LLM confidence weakly tracks logprob signals, with instruction-tuned models showing worse calibration. Gate on verifiable internal signals.

13 min
11 min

Generative vs Cross-Encoder LLM Rerankers: Can Single-Pass Decoding Close the Gap?

7 min

mdlARC's 44% ARC-AGI-1 Claim: What the Small Budget Leaves Out

13 min

FP8 vs MXFP4 vs BF16: Why Your Quantized LLM Disagrees Across GPUs

archive1–50 of 124 · page 1 / 3
  1. Running a 104GB LLM on a 48GB Mac: What Expert Streaming Costs
  2. GPT-6 Astra on ARC-AGI-3: What the Agentic Score Actually Measures
  3. Deepfake KYC Fraud: Tamper-Resilient Watermarks That Recover the Original Face
  4. Detecting AI-Generated Audio: Why Decay Tails Betray Voice Clones
  5. Reward Hacking Starts in the Verifier: Rule Checks vs LLM Judges for Math RL
  6. Can LLMs Train on Their Own Problems? What Zero-Data Self-Play Changes
  7. Why LLM Agent Benchmarks Move When the Harness Changes
  8. Can You Serve LLMs on 2-Bit Weights? What Ultra-Low-Bit Quantization Costs
  9. Do LLMs Still Need BPE? What RL-Trained Tokenizers Change
  10. GLM-5.3-Flash vs Qwen3.8-Flash-Next: Which Budget LLM to Route To
  11. Why Prompt Caching Can Change Model Outputs: A Prefix Invariance Audit
  12. GDPR Deletion Requests vs LLM Weights: What Machine Unlearning Actually Removes
  13. Why Drift Monitors Confuse Covariate Shift With Concept Drift
  14. AI-Generated Apps Look Right, but Do They Actually Work?
  15. Prompt Injection in 3D Scenes: The Attack Surface Multimodal Agents Ignore
  16. Why LLM Log Anomaly Detection Pages You for Nothing
  17. DeepSeek v4 Flash Vision: Routing Images Without Verified Pricing
  18. DeepSeek 32B on RTX 3090: Tokens per Second by Quant and Context
  19. PTXBench: LLMs Can Port GPU Kernels, But Not Beat Tuned Libraries
  20. Can LLMs Reuse Another Model's KV Cache? What Cross-Model Transfer Shows
  21. DeCRIM: Decompose Constraints to Stop Silent Drops in Agent Outputs
  22. Operator-Level Triage for Silent Mixed-Precision Instability
  23. BeyondUncertainty: Weak Confidence Signal for RAG Routing, Not Calibration
  24. Kimi K3 on M1 Max: Bandwidth, Not Capacity, Limits Local MoE Inference
  25. Kimi Linear Cuts KV Cache 75% but Recall Remains the Binding Constraint
  26. Why Chat Leaderboards Do Not Predict Image Quality
  27. Kimi K3 Procurement: Governance Review Over Phantom Government Assessments
  28. DeepSeek Compute Leak: Why Open-Weight Routing Needs a Swap Path
  29. Context Ordering Beats Window Size for Long-Context Agents
  30. Diffusion LLMs: Training Cost, Not Parallel Decoding, Drives Deployment
  31. Open-Weight Routers vs Fable 5: The Routing Math That Actually Matters
  32. DeepSeek-V4 1M Context vs RAG: Why Retrieval Stays
  33. Qwen-Image-3.0 Does Not Exist: Why Self-Hosting Image Models Is Premature
  34. Kimi K3: 2.8T Parameters, MoE Routing, and Self-Hosting Reality
  35. Kimi K3 vs Qwen3.8 Max: Routing Strategy for July 2026
  36. Qwen3.8 Max Release Audit: API, Open Weights, and the License Catch
  37. HuggingFace 100x Inference: Generalizable vs Platform-Locked Optimizations
  38. Kimi K3 Code Arena Rank: Self-Hosting Cost Math for Coding Agents
  39. Can Tool-Adaptive LLM Rerankers Improve RAG Without Always Calling Tools?
  40. Does Speculative Decoding with Progressive Tree Drafting Cut LLM Latency?
  41. Analytic Inference Cuts Bayesian Deep Ensemble Serving Cost, But Leaves Training as the Bottleneck
  42. Tree-of-Thoughts Improves Text-to-Image Prompting by Reasoning Over Hypotheses, Not Pixels
  43. FourierQK's spectral Q/K filter cuts TinyShakespeare loss by 79%, but long-context proof is missing
  44. Tencent Hunyuan 3's Agent Push Has No Public DeepSeek or Qwen Benchmarks Yet
  45. When Does Memory, Not Compute, Decide Who Can Profitably Serve LLMs?
  46. Can We Trust LLM Logic? A Graph-Based Stress Test Finds Three Failure Modes
  47. LongCat-2.0 Hits Claude Opus 4.6 Class on Agents From a 50K-GPU Cluster
  48. When Do Time Series Foundation Models Pay Off? The Break-Even Threshold
  49. VLA Grounder Tests Language Conditioning to Optimize Black-Box Vision-Action Models
  50. Black-Box LLM Architecture Inference: What API Restrictions Reveal About Hidden Model Structure
about this beateditorial framing

Every new foundation model arrives wrapped in a benchmark chart and a press release. This beat exists because the chart almost never tells you what the model actually does — and the press release tells you even less. The interesting work is one layer down: the attention variant that changes the cost curve, the parameterization fix that makes scaling laws actually transfer, the eval protocol whose failure modes flip the standings once you change a decoding parameter or a codec. We cover that layer.

We treat open-weight and closed releases as the same story told from opposite ends of a distribution shift. A trillion-parameter mixture beating a dense competitor on one harness and losing on another is not a contradiction; it is evidence about what the harness measures. Training-efficiency claims, context-window claims, and reasoning-benchmark claims all get the same treatment — read the method section, find the assumption that was load-bearing, and report whether removing it changes the result. When a paper proposes a new safety mechanism, a new compression trick, or a new continual-learning split, we are interested in the part the authors did not want to highlight.

The throughline is comparative and skeptical without being contrarian. Foundation-model research has become the field where the largest gap between published numbers and deployed behavior tends to live. Closing that gap — with reproduction notes, ablation reading, and honest accounting of what generalizes versus what was tuned into the eval — is the beat.