Meta Tagged Employees on Medical and Family Leave for Layoffs Using AI — Leading to a Class-Action Lawsuit

It emerged in court that Meta (formerly Facebook) had used an AI system to automatically tag employees who took medical or family leave as layoff targets. According to a report by The Guardian, a lawsuit filed by former employees alleges that Meta's HR system leveraged AI to lower the performance ratings of employees who took leave and automatically added them to termination lists.

While a growing number of companies are adopting AI for HR decisions, this case looks set to become a significant precedent on the grounds that automated AI classification infringed workers' rights. The outcome of the ruling could have a major impact on future guidelines for corporate use of AI in HR.

Is "Fewer Tokens = Lower Costs" Actually True? Results from a Large-Scale Experiment with Claude Code

A study has been published that examines whether context compression actually leads to cost savings in API-based coding agents (such as Claude Code). The results of a large-scale paired comparison experiment — 2,908 billable runs across 103 tasks, 7 repositories, and 3 models — yielded surprising conclusions.

The three main findings:

Prompt caching accounts for roughly 87% of costs** — even when compression reduces token counts, cache creation and reads remain the dominant cost driver, offsetting the savings

A 38% reduction in tool output actually increased costs by 6.8%** — the confidence intervals also skewed positive, and the correlation between the amount reduced and cost savings is weak (Pearson r = 0.15)

Compression degrades task completion rates** — in SWE-bench-based tests, the patch application success rate worsened from 27/40 to 15/40, because verbatim edit anchors were broken

The study proposes a new evaluation criterion: measure success-adjusted actual billed costs rather than token reduction rates. This is highly practical insight for teams operating RAG pipelines or coding agents.

SRM-LoRA: Mathematically Suppressing LLM Hallucinations with a Riemannian Metric

In a paper accepted at an ICML 2026 Workshop (FoGen), researchers introduced "SRM-LoRA," a method that reduces LLM hallucinations by controlling the direction of LoRA parameter updates using a sub-Riemannian metric.

The method constructs a sensitivity-based Riemannian metric in the LoRA parameter space and suppresses gradients along high-cost update directions (those prone to causing hallucinations). Crucially, it adjusts only the gradient direction during training, **without changing inference-time costs at all**.

Despite being trained only on HaluEval-QA, the model reportedly improved factuality on both related benchmarks and out-of-distribution benchmarks. It also stands out as a demonstration that mathematical theory can genuinely contribute to improving LLMs.

A Multi-Agent Cooperation Benchmark Evaluating 13 LLMs — Cooperation Is the Biggest Bottleneck

A new benchmark called "ALEM" has been released to measure whether populations of LLM agents can cooperate in long-horizon, open-ended environments. When 13 state-of-the-art LLMs were evaluated, the majority of models achieved an average normalized return of only about 6%.

The most interesting finding is that in the hardest setting, **zero-shot Gemini 3.1 Pro performed on par with the best MARL agents trained for one billion steps**. Ablation experiments also revealed that "cooperation" is an independent bottleneck, separate from long-horizon task execution ability, with communication skill having the greatest impact.

The design of inter-agent communication looks set to become key in future multi-agent system development.

Running Qwen 3.5 122B Heretic on Strix Halo — The Community's Challenge

The LocalLLaMA community has released a ROCmFP4 iMatrix quantized version of Qwen 3.5 122B (A10B, i.e., 10B active parameters). On AMD Strix Halo (unified memory), it achieves 60.70 GiB of memory usage with 122B total and 10B active parameters, at a generation speed of 28.45 tok/s.

The community is also actively discussing "what would you run with 384GB (RTX PRO 6000 x4)," exchanging practical insights on local LLM operation for enterprise use cases, such as DeepSeek V4 Flash and running multiple models in parallel.

---