groundy

models & research

80 articles · rss

Top in models & research

models

Can RoboSSM's State-Space Backbone Replace Transformer Imitation Policies?

The RoboSSM preprint swaps the transformer backbone of in-context robot imitation for a Longhorn state-space model and claims LIBERO gains. The full paper is still pending.

jun 19
models

Pruning Experts to Shrink MoE Models: Does Attribution-Guided Compression Beat Magnitude?

Routing architecture decides whether MoE experts can be pruned safely: hard routers keep calibration per expert, soft routers need an aggregate guard.

jun 19
models

GLM-5.2 vs Kimi K2.7 Code: Two Open-Weight Bets on Agentic Coding

GLM-5.2 and Kimi K2.7 Code shipped a day apart as open-weight coding models. One leads every open-weight leaderboard; the other is cheaper and more token-efficient.

models

How Linear Is a Transformer Feed-Forward Block? A New Test Says It's Learned, Not Built In

Per-block linear recoverability in transformer FFNs swings from near-linear to strongly nonlinear between layers, so compression tools must probe each block per checkpoint.

models

GLM-5.2 Benchmarks: What 62.1% SWE-bench Pro and 99.2% AIME Actually Mean

Zhipu published a full benchmark suite for GLM-5.2 on June 19, 2026. Each score targets a different skill domain, and each carries distinct contamination or harness caveats.

models

GLM-5.2 on Terminal-Bench 2.1: Strengths, Gaps, and How to Route Real Coding Tasks

GLM-5.2 scores 81.0 on Terminal-Bench 2.1, trails Claude Opus 4.8 (85.0) on shell tasks, but wins on 1M-token monorepo context. Here is how to route tasks.

models

GLM-5.2 vs Claude Opus 4.8: Open-Weight Coding at Frontier Pricing

GLM-5.2 posts 62.1% on SWE-bench Pro and 81.0 on Terminal-Bench 2.1, four points behind Opus 4.8. MIT weights are self-hostable; flat plan starts at $18/month.

models

GLM-5.2's 753B MoE Costs More to Self-Host Than the MIT License Suggests

GLM-5.2 ships 753B parameters under MIT with strong coding benchmarks, but the full MoE weight load makes self-hosting far heavier than the license terms imply.

more in this beat 1–50 of 72 · page 1 / 2
  1. jun 17 models STAR Replaces Scalar Reward in Text-to-Image RL with Attention-Derived Spatial Maps
  2. jun 15 models Can Editing One Neuron Fix LLM Repetition Loops?
  3. jun 10 models Claude Fable 5 Benchmarks: What FrontierCode, CursorBench, and ViBench Show
  4. jun 10 models Does Attribution Patching Lie? A Fix for a Common Interpretability Shortcut
  5. jun 11 models Can You Make a Multimodal Model Unlearn With Activation Steering?
  6. jun 11 models Why Pruning a Model Can Raise Its Out-of-Distribution Accuracy
  7. jun 09 models Do Unified Multimodal Models Actually Interleave Understanding and Generation?
  8. jun 09 models How LLMs Track Who Did What: The Entity Rebinding Circuit
  9. jun 09 models Claude Fable 5 vs Opus 4.8: When 2x Pricing Is Worth It
  10. jun 09 models Claude Mythos 5 Access Rules: Who Gets Project Glasswing and Why
  11. jun 09 models Fable 5 Distillation Protection: How Anthropic Blocks Model Copying
  12. jun 09 models Skip Fable 5 or Upgrade? When Opus 4.8 and Sonnet 4.6 Are Still Enough
  13. jun 08 models LLM Steganography: Can Defenders Detect Payloads Hidden in Model Output?
  14. jun 08 models Do Privacy Defenses Actually Protect Fine-Tuned LLMs? A New Benchmark
  15. jun 08 models Can You Reconstruct an LLM's System Prompt From Its Activations?
  16. jun 08 models Does Softmax Normalization Limit What Attention Can Represent?
  17. jun 07 models Can an Attacker Steal Your Model's Last Layer From Its Outputs?
  18. jun 06 models Can LLMs Leak Training Data? A New Test Splits Capacity From Intent
  19. jun 06 models When an AI Agent's Tools Break, Can It Recover? A New Benchmark
  20. jun 05 models MiniMax M3 Bets on Sparse Attention for 1M Context. Does the Math Hold?
  21. jun 05 models Can One Model Handle Every CAD Task? UniCAD Tests It
  22. jun 05 models Do Foundation Models Actually Learn Relational Structure In-Context?
  23. jun 05 models Can LLMs Write Better Research Paper Titles Than Authors?
  24. jun 05 models Does Information-Theoretic Example Selection Beat kNN for In-Context Learning?
  25. jun 05 models Do Concept Bottleneck Model Benchmarks Measure Interpretability or Dataset Bias?
  26. jun 05 models Continuous Bit-Width Quantization vs Fixed INT4: Does LiftQuant Beat Discrete?
  27. jun 04 models Federated Learning for Industrial IoT Anomaly Detection: The Data-Locality Tradeoff
  28. jun 04 models Reading Failed LLM Reasoning Traces Won't Tell You Which Ones RL Can Fix
  29. jun 04 models Can You Stitch Two Foundation Models Together Without Retraining?
  30. jun 04 models Do Reasoning LLMs Waste Tokens? OckBench Tries to Measure It
  31. jun 03 models Which Layer Detects LLM Hallucinations Best? The Case Against Fixed-Layer Probes
  32. jun 02 models Cross-Domain RL Training Degrades Capabilities. CARE-RL Reweights to Fix It
  33. jun 02 models LLM Watermarking Without Quality Loss: The Non-Distortionary Approach
  34. jun 01 models Treating LLM Agent Memory as a Database: The VikingMem Approach
  35. jun 01 models Can a Language Model Work Without a Neural Network? A New arXiv Paper Says Yes
  36. jun 01 models Can Code-Generating LLMs Do Engineering Math? FEM-Bench Tests Them
  37. jun 01 models Unlearning Isn't Deletion: arXiv 2505.16831 Shows Machine Unlearning in LLMs Is Reversible
  38. may 31 models Why LLMs Fail at Spatial Reasoning When Planning Navigation
  39. may 31 models Does Giving AI Agents More Skills Help? A Controlled SkillsBench Study
  40. may 30 models Can an LLM Peer-Review Your Paper? A New Behavior Benchmark
  41. may 30 models Anthropic Scaled Sparse Autoencoders to Claude 3 Sonnet. Interpretability Now Costs Compute
  42. may 28 models Tracing Why LLM Agent Memory Fails: A Method for Attributing Errors
  43. may 28 models Persona Prompts Change Who an LLM Recommends as an Expert
  44. may 27 models Opus 4.8 vs Opus 4.7: What Changed and What Did Not
  45. may 27 models Opus 4.8 Batch API: 1M Context, 300k Output, and Team Cost Controls
  46. may 26 models Scale Vectors: Tiny Parameter Subsets That Disproportionately Steer LLM Behavior
  47. may 26 models One Learning Rate Doesn't Fit All: Heavy-Tail Layerwise LR Schedules for LLM Pretraining
  48. may 25 models Audio LLMs Break When the Codec Changes: A Robustness Vector Voice-AI Teams Haven't Tested
  49. may 25 models Do LLMs Know What Not to Say? Causal Evidence for Statistical Preemption
  50. may 24 models Embedding Compression at Training Time: DIVE's Gradient Trick vs Post-Hoc Quantization for Vector DBs
about this beat editorial 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.