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.https://groundy.com/en-ushttps://groundy.com/articles/claude-fable-5-benchmarks-what-frontiercode-cursorbench-and-vibench-show/https://groundy.com/articles/claude-fable-5-benchmarks-what-frontiercode-cursorbench-and-vibench-show/Claude Fable 5 claims top benchmark scores. Verified data shows every model below 14% on FrontierCode Diamond, and partner scores lack public methodology.Sat, 13 Jun 2026 03:20:56 GMTGroundy Editorial2026-06-13T00:00:00.000Zclaude-fable-5frontiercodeai-benchmarkscursorbenchanthropicai-codingGroundy Editorialhttps://groundy.com/articles/does-attribution-patching-lie-a-fix-for-a-common-interpretability-shortcut/https://groundy.com/articles/does-attribution-patching-lie-a-fix-for-a-common-interpretability-shortcut/A June 2026 paper traces attribution patching's errors to downstream non-linearities and proposes a Hessian-vector-product correction that costs one extra backward pass.Sat, 13 Jun 2026 03:20:49 GMTGroundy Editorial2026-06-13T00:00:00.000Zattribution-patchingmechanistic-interpretabilitycircuit-discoveryhessian-correctionllm-interpretabilityactivation-patchingGroundy Editorialhttps://groundy.com/articles/can-you-make-a-multimodal-model-unlearn-with-activation-steering/https://groundy.com/articles/can-you-make-a-multimodal-model-unlearn-with-activation-steering/Steering vectors suppress behavior at runtime without editing weights. Two 2026 papers show they transfer between models, so suppression alone is not unlearning.Sat, 13 Jun 2026 03:20:47 GMTGroundy Editorial2026-06-13T00:00:00.000Zactivation-steeringmachine-unlearningllm-safetysteering-vectorsmodel-evaluationmultimodal-modelsGroundy Editorialhttps://groundy.com/articles/why-pruning-a-model-can-raise-its-out-of-distribution-accuracy/https://groundy.com/articles/why-pruning-a-model-can-raise-its-out-of-distribution-accuracy/Task-aware layer pruning removes distortion-amplifying layers and improves out-of-distribution accuracy, which means standard in-distribution benchmarks miss the real effect.Sat, 13 Jun 2026 03:20:45 GMTGroundy Editorial2026-06-13T00:00:00.000Zmodel-pruningout-of-distributionmodel-evaluationtaopiocageneralizationrepresentation-geometryGroundy Editorialhttps://groundy.com/articles/do-unified-multimodal-models-actually-interleave-understanding-and-generation/https://groundy.com/articles/do-unified-multimodal-models-actually-interleave-understanding-and-generation/IMUG-Bench tests whether unified multimodal models can alternate between understanding and generation in one context, exposing gaps that separate benchmarks conceal.Sat, 13 Jun 2026 03:20:35 GMTGroundy Editorial2026-06-13T00:00:00.000Zmultimodal-modelsbenchmarksimage-generationimage-understandingimug-benchmodel-evaluationGroundy Editorialhttps://groundy.com/articles/how-llms-track-who-did-what-the-entity-rebinding-circuit/https://groundy.com/articles/how-llms-track-who-did-what-the-entity-rebinding-circuit/New research isolates a compact attention-head circuit for entity rebinding in Gemma and Llama, showing tracking failures stem from a binding step, not context length.Wed, 10 Jun 2026 12:05:56 GMTGroundy Editorial2026-06-10T00:00:00.000Zmechanistic-interpretabilityentity-trackingattention-headslong-contextllm-circuitsactivation-patchingGroundy Editorialhttps://groundy.com/articles/claude-fable-5-vs-opus-4-8-when-2x-pricing-is-worth/https://groundy.com/articles/claude-fable-5-vs-opus-4-8-when-2x-pricing-is-worth/Claude Fable 5 prices at $10/$50 per million tokens, 2x Opus 4.8. Frontier research, long-context agents, and molecule design clear the bar. Standard coding does not.Wed, 10 Jun 2026 00:00:00 GMTGroundy Editorial2026-06-10T00:00:00.000Zclaudeanthropicai-pricingfrontier-modelsbenchmarksagentic-codingGroundy Editorialhttps://groundy.com/articles/claude-mythos-5-access-rules-who-gets-project-glasswing-and-why/https://groundy.com/articles/claude-mythos-5-access-rules-who-gets-project-glasswing-and-why/Claude Mythos 5 shares Fable 5's architecture but with safeguards lifted in select areas. Access requires Project Glasswing approval or a biology research designation.Wed, 10 Jun 2026 00:00:00 GMTGroundy Editorial2026-06-10T00:00:00.000Zclaudeanthropicai-safetyfrontier-modelscybersecuritybiology-aiai-policyGroundy Editorialhttps://groundy.com/articles/fable-5-distillation-protection-how-anthropic-blocks-model-copying/https://groundy.com/articles/fable-5-distillation-protection-how-anthropic-blocks-model-copying/Claude Fable 5 ships with distillation protection to prevent capability extraction. A first-principles look at what it is, how it works, and why API consumers should care.Wed, 10 Jun 2026 00:00:00 GMTGroundy Editorial2026-06-10T00:00:00.000Zclaude-fable-5anthropicmodel-securitydistillationai-safetyfrontier-modelsGroundy Editorialhttps://groundy.com/articles/skip-fable-5-or-upgrade-when-opus-4-8-and-sonnet-4-6-are-still-enough/https://groundy.com/articles/skip-fable-5-or-upgrade-when-opus-4-8-and-sonnet-4-6-are-still-enough/Claude Fable 5 costs $10/$50 per MTok, exactly double Opus 4.8. Here is how to decide which tier your workload actually needs and when staying put saves real money.Wed, 10 Jun 2026 00:00:00 GMTGroundy Editorial2026-06-10T00:00:00.000Zclaude-fable-5claude-opus-4-8ai-pricingmodel-selectionanthropicllm-costGroundy Editorialhttps://groundy.com/articles/llm-steganography-can-defenders-detect-payloads-hidden-in-model-output/https://groundy.com/articles/llm-steganography-can-defenders-detect-payloads-hidden-in-model-output/A 2026 proof shows data hidden in LLM output must inflate text complexity. A perplexity proxy catches naive encoders, but adaptive adversaries can evade detection in.Tue, 09 Jun 2026 19:54:36 GMTGroundy Editorial2026-06-09T00:00:00.000Zllm-steganographysteganalysiskolmogorov-complexityperplexityllm-securityoutput-channel-securityGroundy Editorialhttps://groundy.com/articles/do-privacy-defenses-actually-protect-fine-tuned-llms-a-new-benchmark/https://groundy.com/articles/do-privacy-defenses-actually-protect-fine-tuned-llms-a-new-benchmark/A June 2026 benchmark shows passing privacy attack probes on fine-tuned LLMs is not a formal guarantee, exposing a compliance gap for teams deploying models on customer data.Tue, 09 Jun 2026 13:09:48 GMTGroundy Editorial2026-06-09T00:00:00.000Zllm-privacyfine-tuningdifferential-privacymembership-inferencemodel-securitycomplianceGroundy Editorialhttps://groundy.com/articles/can-you-reconstruct-an-llms-system-prompt-from-its-activations/https://groundy.com/articles/can-you-reconstruct-an-llms-system-prompt-from-its-activations/PRISM recovers full instruction sets inside frozen LLMs from hidden states, enabling anyone with activation access to reconstruct system prompts without output probing.Tue, 09 Jun 2026 12:45:38 GMTGroundy Editorial2026-06-09T00:00:00.000Zllm-interpretabilityactivation-probessystem-prompt-extractionprismmodel-securityai-safetyGroundy Editorialhttps://groundy.com/articles/does-softmax-normalization-limit-what-attention-can-represent/https://groundy.com/articles/does-softmax-normalization-limit-what-attention-can-represent/A new paper proves softmax normalization imposes geometric separation bounds on token vectors, constraining what attention can represent as context length grows.Tue, 09 Jun 2026 00:55:16 GMTGroundy Editorial2026-06-09T00:00:00.000Zsoftmax-attentionattention-mechanismtransformer-architecturenormalizationneural-network-theorymodel-architectureGroundy Editorialhttps://groundy.com/articles/can-an-attacker-steal-your-models-last-layer-from-its-outputs/https://groundy.com/articles/can-an-attacker-steal-your-models-last-layer-from-its-outputs/A new geometric proof shows API outputs alone suffice to recover a transformer's final projection matrix up to a rotation, while deeper layers are provably irrecoverable.Mon, 08 Jun 2026 19:37:14 GMTGroundy Editorial2026-06-08T00:00:00.000Zmodel-stealingmodel-securityllm-apisai-safetytransformer-architecturedifferential-privacyGroundy Editorialhttps://groundy.com/articles/can-llms-leak-training-data-a-new-test-splits-capacity-from-intent/https://groundy.com/articles/can-llms-leak-training-data-a-new-test-splits-capacity-from-intent/PropMe splits memorization audits into capability and propensity, showing that single-metric leakage reports understate what targeted prompts can extract from LLMs.Sun, 07 Jun 2026 12:55:22 GMTGroundy Editorial2026-06-07T00:00:00.000Zllm-memorizationgdpr-compliancemodel-evaluationdata-extractiontraining-dataai-safetyGroundy Editorialhttps://groundy.com/articles/when-an-ai-agents-tools-break-can-it-recover-a-new-benchmark/https://groundy.com/articles/when-an-ai-agents-tools-break-can-it-recover-a-new-benchmark/ToolMaze, a new arXiv benchmark, shows LLM agents' recovery rates drop 37% when tools return corrupted data, exposing a gap in how agent reliability is measured.Sun, 07 Jun 2026 05:32:08 GMTGroundy Editorial2026-06-07T00:00:00.000Zllm-agentstool-failurebenchmarkagent-reliabilityfault-tolerancedynamic-replanningGroundy Editorialhttps://groundy.com/articles/minimax-m3-bets-on-sparse-attention-for-1m-context-does-the-math-hold/https://groundy.com/articles/minimax-m3-bets-on-sparse-attention-for-1m-context-does-the-math-hold/MiniMax claims M3 handles 1M tokens via sparse attention, but published no technical report or independent benchmarks. Retrieval quality at full context is unverified.Sat, 06 Jun 2026 08:46:35 GMTGroundy Editorial2026-06-06T00:00:00.000Zsparse-attentionminimax-m3long-contextllm-benchmarksinference-costretrieval-qualityGroundy Editorialhttps://groundy.com/articles/can-one-model-handle-every-cad-task-unicad-tests/https://groundy.com/articles/can-one-model-handle-every-cad-task-unicad-tests/UniCAD introduces a unified benchmark and single multi-modal model for CAD reconstruction, generation, and question answering, challenging the field's per-task silos.Sat, 06 Jun 2026 08:00:51 GMTGroundy Editorial2026-06-06T00:00:00.000Zcadmulti-modal-modelsdeep-learningbenchmarks3d-reconstructiongenerative-designGroundy Editorialhttps://groundy.com/articles/do-foundation-models-actually-learn-relational-structure-in-context/https://groundy.com/articles/do-foundation-models-actually-learn-relational-structure-in-context/OpenRFM shows relational in-context learning collapses on sparse joins and introduces a dual-stage architecture that surpasses the commercial KumoRFMv1 baseline.Sat, 06 Jun 2026 06:39:42 GMTGroundy Editorial2026-06-06T00:00:00.000Zrelational-foundation-modelsin-context-learningtabular-modelsopenrfmrelational-learningpre-trainingGroundy Editorialhttps://groundy.com/articles/can-llms-write-better-research-paper-titles-than-authors/https://groundy.com/articles/can-llms-write-better-research-paper-titles-than-authors/A new study claims LLMs write 'appropriate' research titles, but the evidence rests on similarity metrics that measure pattern matching, not whether titles actually serve.Sat, 06 Jun 2026 05:36:20 GMTGroundy Editorial2026-06-06T00:00:00.000Zllm-evaluationresearch-titlesacademic-publishingtext-metricspegasusarxivGroundy Editorialhttps://groundy.com/articles/does-information-theoretic-example-selection-beat-knn-for-in-context-learning/https://groundy.com/articles/does-information-theoretic-example-selection-beat-knn-for-in-context-learning/KITE swaps cosine-similarity kNN for a kernelized information-theoretic selector in few-shot prompting, reporting classification gains but adding inference compute overhead.Sat, 06 Jun 2026 05:21:48 GMTGroundy Editorial2026-06-06T00:00:00.000Zin-context-learningfew-shot-learningragexample-selectionkernel-methodsretrieval-optimizationclassificationGroundy Editorialhttps://groundy.com/articles/do-concept-bottleneck-model-benchmarks-measure-interpretability-or-dataset-bias/https://groundy.com/articles/do-concept-bottleneck-model-benchmarks-measure-interpretability-or-dataset-bias/Standard concept bottleneck model benchmarks confound genuine concept learning with dataset shortcuts. Synthetic benchmarks from Skirzynski et al. expose the gap.Sat, 06 Jun 2026 03:27:18 GMTGroundy Editorial2026-06-06T00:00:00.000Zconcept-bottleneck-modelsinterpretabilitysynthetic-benchmarksinformation-leakagemodel-evaluationconfoundingGroundy Editorialhttps://groundy.com/articles/continuous-bit-width-quantization-vs-fixed-int4-does-liftquant-beat-discrete/https://groundy.com/articles/continuous-bit-width-quantization-vs-fixed-int4-does-liftquant-beat-discrete/LiftQuant replaces 2/4/8-bit quantization with continuous bit-width via dimensional lifting. A 70B model at 2.4 bits fits 24GB. Kernel support is the bottleneck.Sat, 06 Jun 2026 01:26:39 GMTGroundy Editorial2026-06-06T00:00:00.000Zquantizationllm-inferenceliftquantmixed-precisionmodel-compressionsub-4bit-quantizationGroundy Editorialhttps://groundy.com/articles/federated-learning-for-industrial-iot-anomaly-detection-the-data-locality/https://groundy.com/articles/federated-learning-for-industrial-iot-anomaly-detection-the-data-locality/A DEXA 2026 paper proposes a cyclic-dynamics benchmark for federated anomaly detection, exposing the gap between on-site compliance gains and unknown convergence costs.Fri, 05 Jun 2026 23:29:18 GMTGroundy Editorial2026-06-05T00:00:00.000Zfederated-learninganomaly-detectionindustrial-iottime-seriesdata-localitycyclic-dynamicsGroundy Editorialhttps://groundy.com/articles/reading-failed-llm-reasoning-traces-wont-tell-you-which-ones-rl-can-fix/https://groundy.com/articles/reading-failed-llm-reasoning-traces-wont-tell-you-which-ones-rl-can-fix/A new preprint finds the fixability of failed LLM reasoning rollouts under RL is predictable from distributional statistics, not from reading chain-of-thought text.Fri, 05 Jun 2026 18:08:14 GMTGroundy Editorial2026-06-05T00:00:00.000Zreasoning-rlchain-of-thoughtreinforcement-learningpost-trainingprocess-reward-modelsllm-reasoningtest-time-computeGroundy Editorialhttps://groundy.com/articles/can-you-stitch-two-foundation-models-together-without-retraining/https://groundy.com/articles/can-you-stitch-two-foundation-models-together-without-retraining/Splicing layers from independently trained foundation models fails without targeted training at the join point. A two-stage recipe called Final Feature Matching makes it work.Fri, 05 Jun 2026 16:44:35 GMTGroundy Editorial2026-06-05T00:00:00.000Zmodel-stitchingfoundation-modelsvision-modelsmodel-mergingtransfer-learningrepresentation-learningGroundy Editorialhttps://groundy.com/articles/do-reasoning-llms-waste-tokens-ockbench-tries-to-measure/https://groundy.com/articles/do-reasoning-llms-waste-tokens-ockbench-tries-to-measure/OckBench scores 37 reasoning LLMs on token efficiency alongside accuracy, finding comparably accurate models differ by up to 26× in token cost under per-token billing.Fri, 05 Jun 2026 08:44:51 GMTGroundy Editorial2026-06-10T00:00:00.000Zockbenchllm-reasoningtoken-efficiencymodel-benchmarkinginference-costreasoning-modelsGroundy Editorialhttps://groundy.com/articles/which-layer-detects-llm-hallucinations-best-the-case-against-fixed-layer-probes/https://groundy.com/articles/which-layer-detects-llm-hallucinations-best-the-case-against-fixed-layer-probes/An ICML 2026 paper finds that fixed-layer hallucination probes miss detection signal, and proposes FEPoID, a training-free method to calibrate layer choice per model.Thu, 04 Jun 2026 10:09:05 GMTGroundy Editorial2026-06-04T00:00:00.000Zhallucination-detectionhidden-statesllm-probesintrinsic-dimensiontransformer-layersicml-2026Groundy Editorialhttps://groundy.com/articles/cross-domain-rl-training-degrades-capabilities-care-rl-reweights-to-fix/https://groundy.com/articles/cross-domain-rl-training-degrades-capabilities-care-rl-reweights-to-fix/CARE-RL shows that pooling math, code, and chat into one RL run causes silent capability erosion across domains, and proposes gradient subspace projection to reweight updates.Wed, 03 Jun 2026 18:17:35 GMTGroundy Editorial2026-06-03T00:00:00.000Zreinforcement-learningmulti-domain-trainingcapability-interferencegradient-editingllm-post-trainingreward-signal-designGroundy Editorialhttps://groundy.com/articles/llm-watermarking-without-quality-loss-the-non-distortionary-approach/https://groundy.com/articles/llm-watermarking-without-quality-loss-the-non-distortionary-approach/LUNA's POS-adaptive watermark claims AUROC 0.9959 with 0.045 perplexity shift across six languages, but paraphrase robustness remains untested for all distortion-free schemes.Wed, 03 Jun 2026 14:51:29 GMTGroundy Editorial2026-06-03T00:00:00.000Zllm-watermarkingai-content-detectionprovenance-trackingtext-generationnlp-researchmultilingual-nlpGroundy Editorialhttps://groundy.com/articles/treating-llm-agent-memory-as-a-database-the-vikingmem-approach/https://groundy.com/articles/treating-llm-agent-memory-as-a-database-the-vikingmem-approach/VikingMem treats LLM agent memory as a database with events, entities, and temporal compression, reporting up to 30% better retrieval than current approaches.Tue, 02 Jun 2026 16:32:05 GMTGroundy Editorial2026-06-02T00:00:00.000Zllm-agentsagent-memoryvikingmemvector-databasesmemory-managementvldb-2026Groundy Editorialhttps://groundy.com/articles/can-a-language-model-work-without-a-neural-network-a-new-arxiv-paper-says-yes/https://groundy.com/articles/can-a-language-model-work-without-a-neural-network-a-new-arxiv-paper-says-yes/A single-author arXiv preprint claims an RBF-network variant can build a language model without backpropagation-trained deep nets, solving for global loss optimum in one step.Tue, 02 Jun 2026 15:03:29 GMTGroundy Editorial2026-06-02T00:00:00.000Zrbf-networkslanguage-modelstransformer-alternativesarxivmachine-learning-architecturebackpropagationGroundy Editorialhttps://groundy.com/articles/can-code-generating-llms-do-engineering-math-fem-bench-tests-them/https://groundy.com/articles/can-code-generating-llms-do-engineering-math-fem-bench-tests-them/FEM-Bench tests 33 finite element tasks and finds Gemini 3 Pro solved 30 in five attempts. The risk: LLM solvers that compile and run but return physically wrong results.Tue, 02 Jun 2026 14:22:58 GMTGroundy Editorial2026-06-02T00:00:00.000Zfem-benchcode-generationllm-benchmarksfinite-element-methodscientific-computingllm-evaluationGroundy Editorialhttps://groundy.com/articles/unlearning-isnt-deletion-arxiv-2505-16831-shows-machine-unlearning-in-llms/https://groundy.com/articles/unlearning-isnt-deletion-arxiv-2505-16831-shows-machine-unlearning-in-llms/Two independent studies confirm machine unlearning methods suppress outputs without erasing internal representations, making GDPR compliance claims unverifiable.Tue, 02 Jun 2026 12:36:09 GMTGroundy Editorial2026-06-02T00:00:00.000Zmachine-unlearningllm-securitygdpr-compliancemodel-representationai-safetydata-privacyGroundy Editorialhttps://groundy.com/articles/why-llms-fail-at-spatial-reasoning-when-planning-navigation/https://groundy.com/articles/why-llms-fail-at-spatial-reasoning-when-planning-navigation/LLMs fail at spatial navigation because training text encodes geometry poorly. Two papers show explicit structural scaffolding, not prompt tweaks, is the fix teams need.Mon, 01 Jun 2026 21:05:13 GMTGroundy Editorial2026-06-02T00:00:00.000Zspatial-reasoningllm-navigationinductive-biasembodied-agentssearch-treestraining-data-biasGroundy Editorialhttps://groundy.com/articles/does-giving-ai-agents-more-skills-help-a-controlled-skillsbench-study/https://groundy.com/articles/does-giving-ai-agents-more-skills-help-a-controlled-skillsbench-study/SkillsBench study: curated skills lift agent pass rates 18 to 36 pp, while documentation granularity shifts outcomes under 1 pp. Curation, not polish, is the bottleneck.Mon, 01 Jun 2026 10:49:42 GMTGroundy Editorial2026-06-02T00:00:00.000Zagent-skillsllm-benchmarksskillsbenchai-agentsmodel-evaluationskill-catalogsGroundy Editorialhttps://groundy.com/articles/can-an-llm-peer-review-your-paper-a-new-behavior-benchmark/https://groundy.com/articles/can-an-llm-peer-review-your-paper-a-new-behavior-benchmark/PRAIB benchmarks LLM-generated peer reviews across 11,000 reviews on 1,000 papers, finding positive bias, compressed variance, and systematically overlooked weaknesses.Sun, 31 May 2026 13:26:57 GMTGroundy Editorial2026-06-02T00:00:00.000Zpeer-reviewllm-benchmarkai-peer-reviewresearch-integrityconference-reviewllm-biasGroundy Editorialhttps://groundy.com/articles/anthropic-scaled-sparse-autoencoders-to-claude-3-sonnet-interpretability-now/https://groundy.com/articles/anthropic-scaled-sparse-autoencoders-to-claude-3-sonnet-interpretability-now/Anthropic extracted 34M interpretable features from Claude 3 Sonnet, proving sparse autoencoders work on production models. Interpretability now has its own compute budget.Sun, 31 May 2026 10:23:06 GMTGroundy Editorial2026-06-10T00:00:00.000Zsparse-autoencodersmechanistic-interpretabilityclaude-3-sonnetai-safetydictionary-learninganthropicGroundy Editorialhttps://groundy.com/articles/tracing-why-llm-agent-memory-fails-a-method-for-attributing-errors/https://groundy.com/articles/tracing-why-llm-agent-memory-fails-a-method-for-attributing-errors/MemTrace constructs provenance graphs across every memory operation in an LLM agent, tracing wrong answers to the exact operation that corrupted state across sessions.Fri, 29 May 2026 17:30:46 GMTGroundy Editorial2026-05-29T00:00:00.000Zllm-memorydebuggingragagent-frameworkserror-attributionprovenanceGroundy Editorialhttps://groundy.com/articles/persona-prompts-change-who-an-llm-recommends-as-an-expert/https://groundy.com/articles/persona-prompts-change-who-an-llm-recommends-as-an-expert/A 43-model audit finds that geographic and role framing in LLM prompts systematically shifts which scholars get recommended as experts, with no neutral default.Fri, 29 May 2026 15:41:56 GMTGroundy Editorial2026-05-29T00:00:00.000Zllm-biaspersona-promptsexpert-recommendationscholar-discoveryai-fairnessrecommendation-systemsGroundy Editorialhttps://groundy.com/articles/opus-4-8-vs-opus-4-7-what-changed-and-what-did-not/https://groundy.com/articles/opus-4-8-vs-opus-4-7-what-changed-and-what-did-not/Anthropic's Opus 4.8 raises SWE-Bench Pro from 64.3% to 69.2% and cuts code-flaw pass-through fourfold at unchanged $5/$25 pricing. A fast mode at $10/$50 runs 2.5x quicker.Thu, 28 May 2026 13:30:41 GMTGroundy Editorial2026-06-10T00:00:00.000Zclaudeanthropicopus-48model-releasebenchmarksagentic-codingGroundy Editorialhttps://groundy.com/articles/opus-4-8-batch-api-1m-context-300k-output-and-team-cost-controls/https://groundy.com/articles/opus-4-8-batch-api-1m-context-300k-output-and-team-cost-controls/Opus 4.8 has a 1M token context window (200k on Foundry), 128k standard output, and 300k output via Batch API beta. January 2026 cutoff. Batch design and quota allocation.Thu, 28 May 2026 12:14:25 GMTGroundy Editorial2026-06-10T00:00:00.000Zclaude-opusbatch-apianthropicrate-limitscontext-windowmodel-releaseteam-infrastructureGroundy Editorialhttps://groundy.com/articles/scale-vectors-tiny-parameter-subsets-that-disproportionately-steer-llm-behavior/https://groundy.com/articles/scale-vectors-tiny-parameter-subsets-that-disproportionately-steer-llm-behavior/Scale vectors are a negligible parameter class in LLM normalization layers whose outsized optimization role makes them high-value targets for quantization and safety editing.Wed, 27 May 2026 19:21:51 GMTGroundy Editorial2026-05-28T00:00:00.000Zscale-vectorsllm-quantizationmechanistic-interpretabilitymodel-normalizationmodel-compressionllm-trainingGroundy Editorialhttps://groundy.com/articles/one-learning-rate-doesnt-fit-all-heavy-tail-layerwise-lr-schedules-for-llm/https://groundy.com/articles/one-learning-rate-doesnt-fit-all-heavy-tail-layerwise-lr-schedules-for-llm/LLR assigns per-layer learning rates from spectral heavy-tail diagnostics during LLM pretraining, achieving 1.5x faster convergence and up to 2 pp higher zero-shot accuracy.Wed, 27 May 2026 14:01:42 GMTGroundy Editorial2026-05-27T00:00:00.000Zllm-pretraininglearning-ratespectral-analysisoptimizertransformer-trainingicml-2026Groundy Editorialhttps://groundy.com/articles/audio-llms-break-when-the-codec-changes-a-robustness-vector-voice-ai-teams/https://groundy.com/articles/audio-llms-break-when-the-codec-changes-a-robustness-vector-voice-ai-teams/CodecAttack achieves 85.5% attack success on audio LLMs by optimizing in codec latent space, with 100% zero-shot transfer to MP3, proving lossy compression fails as a defense.Tue, 26 May 2026 11:55:22 GMTGroundy Editorial2026-05-26T00:00:00.000Zadversarial-audioaudio-llmscodec-robustnessvoice-aiadversarial-mlaudio-securityGroundy Editorialhttps://groundy.com/articles/do-llms-know-what-not-to-say-causal-evidence-for-statistical-preemption/https://groundy.com/articles/do-llms-know-what-not-to-say-causal-evidence-for-statistical-preemption/New causal evidence shows LLMs suppress wrong continuations during pretraining via statistical preemption, suggesting output-layer safety fixes may target the wrong layer.Tue, 26 May 2026 10:04:15 GMTGroundy Editorial2026-05-26T00:00:00.000Zstatistical-preemptionllm-safetymodel-interpretabilityhallucinationcausal-probingpretrainingGroundy Editorialhttps://groundy.com/articles/embedding-compression-at-training-time-dives-gradient-trick-vs-post-hoc/https://groundy.com/articles/embedding-compression-at-training-time-dives-gradient-trick-vs-post-hoc/DIVE's gradient-limited adapter outperforms baselines for embedding compression, but training-time methods lock RAG pipelines to specific adapters and raise refresh costs.Mon, 25 May 2026 20:33:59 GMTGroundy Editorial2026-05-26T00:00:00.000Zembedding-compressionragvector-databasesdiveadapter-methodsGroundy Editorialhttps://groundy.com/articles/p-hyperparameter-transfer-has-an-embedding-layer-hole-new-arxiv-paper-says/https://groundy.com/articles/p-hyperparameter-transfer-has-an-embedding-layer-hole-new-arxiv-paper-says/An arXiv paper shows the embedding learning rate accounts for most of μP's advantage over standard parameterization, and a single scaling fix recovers the bulk of the benefit.Mon, 25 May 2026 20:07:07 GMTGroundy Editorial2026-05-26T00:00:00.000Zmuphyperparameter-transferembedding-layeradamwmodel-scalingtraining-optimizationGroundy Editorialhttps://groundy.com/articles/project-glasswing-one-month-in-ai-bug-discovery-has-outpaced-the-patch-pipeline/https://groundy.com/articles/project-glasswing-one-month-in-ai-bug-discovery-has-outpaced-the-patch-pipeline/Anthropic's Glasswing found over 10,000 high-severity vulnerabilities in one month. Only 97 are patched. The bottleneck shifted from discovery to triage, and it is structural.Sun, 24 May 2026 21:12:14 GMTGroundy Editorial2026-06-10T00:00:00.000Zai-securityvulnerability-disclosureanthropicclaude-mythoscybersecurityinterpretabilityGroundy Editorial