The archive · Page 2 of 6
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.
25–48 of 139 articles · Newest first
arXiv 2508.06753 reports 2-bit LLM serving with up to 7x speedups on Intel Xe2. For self-hosters, the binding cost is not memory but task-specific eval coverage to verify QAT-
An ICML 2026 paper shows RL can learn token boundaries end-to-end, beating straight-through baselines at 100M parameters. For serving or fine-tuning, keep your frozen BPE or S
GLM-5.3-Flash and Qwen3.8-Flash-Next lack confirmed pricing and independent evals. Hold your router, verify per-token costs, and run a 50-case domain holdout before switching.
A new arXiv audit shows attention masks miss causality leaks in state-space models. Run a two-pass check before trusting prefix caches on hybrid stacks.
GDPR erasure demands hit LLM weights, but retraining is the only defensible fix. Behavior-level unlearning passes probes without proving data is gone, leaving compliance teams
A new preprint proposes CJSD, a two-discriminator test that separates covariate shift from concept drift. This distinction determines whether to retrain or recalibrate, fixing
MobileForge shows AI apps compile but fail navigation. Shift review from screenshots to project-level state tests to ship reliable generated frontends.
arXiv 2602.07104 shows 3D object placement injects instructions into multimodal agents, bypassing text and image filters. Teams must gate modalities and scope actions to.
arXiv 2608.17965 shows LLM log detectors are overconfident in wrong verdicts. Gate paging on calibrated confidence, not raw F1, to prevent alert fatigue and ignored critical.
DeepSeek v4-flash-vision-exp lacks verified pricing and benchmarks. This guide prices Qwen-VL alternatives and outlines a fallback protocol for experimental vision endpoints.
Bandwidth math predicts 32-39 tok/s for Q4_K_M DeepSeek 32B on an RTX 3090 at 8k context. 32k context overflows 24GB VRAM. No measured benchmarks exist, so every figure is.
PTXBench shows LLMs can generate architecture-specific PTX for H100 and B200, but no model matches frontier libraries. Use it to cut hot-loop porting costs, not to replace.
A new preprint proposes a lightweight reader to reuse KV caches across models, but evidence suggests caches remain model-bound. Learn when recompute beats adaptation.
DeCRIM shows that decomposing multi-constraint instructions into individually checkable units reduces silent drops by 7-8% on benchmarks, shifting reliability work from.
arXiv 2607.25494 introduces a single-pass tool to localize silent numerical instability in bf16 and fp16 training. Treat it as a triage layer for unexplained loss spikes, not.
arXiv:2607.25600 shows verbalized confidence is a weak but real routing signal for RAG. It saves 20.4% retrieval calls for 28.2% token overhead. The probe is poorly.
Running Kimi K3 on an M1 Max proves local MoE feasibility via expert offloading, but unified memory bandwidth caps throughput. Treat this as a prototyping tool, not a serving.
Kimi Linear cuts KV cache 75% and boosts decode 6x for 1M-token contexts, but recall stays the binding constraint. Route memory-bound jobs only after benchmarking in-context.
A July 22 visual test comparing GPT-5.6, Claude, Gemini, and Grok highlights a structural gap: chat models lack dedicated image backends. Teams must route by subtask using.
Moonshot AI's Kimi K3 release triggers governance review for regulated teams. Beijing jurisdiction, Anthropic accusations, and missing government assessment require internal.
A leaked transcript claims DeepSeek faces a compute gap, but V4 shipped in April. Teams should treat this as a resilience prompt to pin Qwen as a swap-in spine, not panic.
ARBIGRAPH shows tool agents lose 33.3% accuracy on dependent chains despite fitting context. Context ordering, truncation, and caching now beat raw window size for agent.
Parallel decoding has not cut serving costs for diffusion LLMs. arXiv:2605.13026 closes training gaps by 4x, but KV-cache absence keeps inference slower than autoregressive.
Fable 5 pricing and caching set the bar for open-weight routers. The Echo claim lacks verification. Routing wins only for high-volume, cache-unfriendly routine traffic.