The week in AI & systems

Top of the week

1. Models & trainingPaperWed, Oct 7

   LoGRA cuts LLM RL training memory by up to 45.7%

   LoGRA stores learning signals as low-rank gradient sketches and uses predicted-KL step control; it reports 45.7% lower average memory and stable 27B-model training for over 1,100 steps on one eight-GPU node. This could make RL post-training feasible where dense Adam runs out of memory.

2. Models & trainingPaperTue, Oct 6

   Fixed-point training improves looped models and speeds prefill and RL updates

   The work uses learned depth priors and orthogonal input injection to improve perplexity across 100M–1.6B models, and reports 1.79× faster student prefill and 2× faster RL updates. Its fixed-point analysis connects model architecture choices to training and inference costs.

3. AgentsPaperSun, Oct 11

   Agent-skill copy networks reveal where security fixes fail to propagate

   A dated network covering 2,193,119 skill adoptions finds that copied skills rarely change with their source; auditing the top 100 ranked repositories prevents 14.9% of later high-risk adoptions, versus 0.5% for the 100 most-starred. The findings offer concrete ways to prioritize agent-skill supply-chain reviews.

4. Cloud & infraPaperWed, Oct 7

   Perseus found 304 fail-slow cases across 248,000 drives and cut tail latency by 48%

   Perseus uses a regression-based detector to identify drive-level fail-slow behavior; isolating detected drives reduced node-level 99.99th-percentile tail latency by 48%. The authors also released a production-derived dataset for studying storage degradation.

5. AgentsPaperTue, Oct 6

   OSWorld-Pro evaluates computer-use agents across subgoals and action failures

   OSWorld-Pro provides over 300 tasks, 2,800 subgoals, and 67,000 human annotations; its top model scores 75.7% versus 83.4% on OSWorld. Process-level evaluation can expose whether agents fail on inputs, clicks, or other steps rather than only judging final deliverables.

6. ArchitecturePaperFri, Oct 9

   SparseEngine reports over 10× throughput with KV eviction

   SparseEngine gives sparse attention methods a shared lifecycle contract and adds cross-request cache management; it reports over 10× throughput with KV eviction and over 2× end-to-end speedup on agent benchmarks. Its serving architecture and results are directly relevant to long-context inference systems.

7. SecurityVulnerabilityTue, Oct 6

   Cloudflare Containers vulnerability exposed residual disk blocks across tenants

   Researchers showed that a customer could recover residual disk blocks previously used by Containers on the same host, though they could not target a particular customer or workload. The cross-tenant isolation failure is directly relevant to engineers operating multi-tenant infrastructure.

8. InferenceBenchmarkMon, Oct 5

   Qwen 125B runs locally at up to 94 tokens/s on tested consumer GPUs

   Strata reports generation speeds of 53–94 tokens/s across quantizations on an RTX 5070 and an RX 9070 XT, with prompt processing also measured. Engineers weighing local inference can compare the speed and hardware trade-offs against server-based deployment.

9. Models & trainingPaperTue, Oct 6

   Dust uses token-level activation perturbations to train transformers without backpropagation

   Dust perturbs activations independently at each token; the authors report competitive results against backprop and estimate 1,000–10,000× greater efficiency than EGGROLL from one million tokens onward. The method and scaling results make it worth watching for teams exploring alternatives to standard training.

10. ResearchPaperThu, Oct 8

    Lean verifies an optimal packing of 11 squares with no admissions

    The verification run accepted 7,920 Lean modules and reports zero admissions; selected exact numerical checks use native_decide, so the claim is not kernel-only. The proof and explicit trust boundary make this a concrete reference for formal verification workflows.

Also worth your time