Microsoft researcher Dimitris Papailiopoulos leveraged GPT-5.6 and Fable 5 to prove that 「LMMSE rounding + greedy bit-flipping」 can achieve exact recovery in MIMO detection at the 2logN maximum likelihood threshold. The algorithm requires O(NlogN) steps and total computational complexity of O(N³), transforming exponential brute-force search into polynomial time. The two models contributed AMP preliminary proof and symbolic LMMSE derivation respectively; the researcher cross-simplified and manually verified each line over seven days. Results also show that below this threshold, maximum likelihood detection itself fails.

The United Nations has established its first independent international AI scientific panel, structured after the IPCC in climate science, aiming to consolidate evidence on AI capabilities and risks and form a global technical consensus for member states' governance negotiations. This mechanism separates scientific assessment from political deliberation within the 「Global AI Governance Dialogue」, and promotes standardized Capability Evals using frameworks like Inspect, covering risks such as pathogen synthesis and deceptive reasoning. The panel also plans to provide ongoing evaluation support for member states lacking budget or technical capacity to establish national AI safety institutions.

A case summary on corporate AI spending reveals that Amazon spent $1.8 million on a task to complete author information—860% over budget—which went unnoticed for five months and was ultimately never deployed. Meta's Claude usage leaderboard for 85,000 employees implies a monthly bill of approximately $221 million based on estimates in the report. The material also states that only 26% of enterprises have full visibility into their AI expenditures; after Uber exhausted its annual AI coding budget in four months, it set a monthly cap of $1,500 per employee, highlighting cost risks from token-based incentives without output tracking, retry controls, or centralized monitoring metrics.

TarantuBench-v2 releases 10,000 AI-generated vulnerable web applications for training and evaluating model cybersecurity capabilities. The benchmark no longer only checks whether the flag is captured, but also verifies application-layer exploitation milestones and HTTP traces to detect reward hacking where models bypass intended paths or attack evaluation infrastructure. The authors argue that open security tools benefit defenders, but explicitly acknowledge that all data are synthetic, each application contains only a single vulnerability, correlation studies with other security benchmarks remain incomplete, and cross-benchmark transfer effects after fine-tuning have not been validated.

CalibForge introduces an adversarial multi-solver calibration framework for terminal agent training data, leveraging solver disagreements and relationships where 「strong models succeed while weak models fail」 to keep synthetic tasks within a learnable difficulty range. It generates 5,431 terminal tasks; according to the project, trained models achieve a 24.71-point improvement on Terminal-Bench 2.0, 27.68 points on SWE-bench Pro, and 30.04 points on Doc2Repo relative to baselines. Cross-benchmark gains suggest enhanced general environment operation capability, though results require independent replication.

Intel has open-sourced LLM Scaler, optimized for text, image, and video generation on Arc Pro B60 and B70 GPUs, supporting models such as DeepSeek and Qwen, and offering INT4, FP8, and MXFP4 quantization paths. The project integrates with vLLM, SGLang, and ComfyUI, enabling developers to reuse existing inference and generative AI workflows when deploying various models on Intel graphics cards. Current materials do not disclose specific throughput, latency, or performance relative to competitors, so claims of 「best performance」 remain aspirational; the code repository is publicly available for assessing compatibility and real-world benefits.

The OOPSLA 2026 project 「just-tri-it」 proposes semantic triangulation, using independently transformed equivalent problems to cross-check LLM-generated programs and reduce code hallucination. The study argues that repeated sampling often produces correlated errors, and simple majority voting may amplify identical mistakes; tests or formal specifications generated by the same model may inherit similar biases. This method requires problem transformations that trigger different algorithms and uses bijective mappings to expose contradictions more easily. The paper proves under its assumed model that triangular consistency is more reliable, and related code and materials are publicly available on GitHub.

The Model DNA method determines LLM lineage using three public signals: model configuration, tokenizer overlap, and Linear CKA in embedding space. A single matching field is insufficient evidence; alignment across five or more indicators forms a strong fingerprint. CKA is robust against rotation and scaling, making it harder to evade than direct cosine similarity through representation transformation. Layer-wise CKA, shared token counts, and confidence intervals can further enhance resolution. However, the method still struggles to distinguish derivative models after large-scale continued pretraining, and thresholds vary with candidate sets, so it can only suggest probabilistic lineage—not compliance status or developer intent.