Semantic review is behind discovery. Evidence and Pulse may be incomplete until the backlog is cleared.
2 material first-party/frontier/open-problem candidate(s) have waited more than 6 hours for semantic review.
EVIDENCE REGISTER
A chronological register of published signals. Source class, verification state, evidence quality and relevance remain visible before interpretation.
Google launched agentic video understanding for Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite. The system dynamically searches and reinspects video segments across frames, audio and transcripts; Google reports up to 88% lower token use, up to 66% lower analysis cost and up to 7% higher accuracy on tested video-analysis benchmarks.
Anthropic released Claude Fable 5.1, a generally available Mythos-class model for demanding reasoning, coding, research and long-running agentic work. Anthropic reports 52.6% on Terminal-Bench-Science 0.1 versus 24.7% for Fable 5 and 55.8% on Terminal-Bench 4.0 versus 42.0% for Fable 5. The model has a 1M-token context window and 128K maximum output. Release and availability are independently corroborated, while the headline capability benchmarks remain primarily provider-reported and require independent reproduction.
James Freitag and Scott Mutchnik report a counterexample to the stable forking conjecture, a long-standing problem in model theory discussed since 1996. The authors state in the abstract that they found the counterexample using ChatGPT 5.6. The result is currently an arXiv preprint: the theorem and AI role are primary-source confirmed, but no peer review or independent mathematical reproduction was identified in this pass.
PaperGym converts scientific papers into research-planning environments with separated questions and grading criteria, then combines rubric-conditioned self-distillation with rubric-reward training. Across Qwen3 models from 1.7B to 8B, the authors report five-benchmark average gains of 4.8 to 5.6 points and release the 20,000-instance corpus, benchmarks, models and training code.
ASPIRE evaluates whether agents can turn vague goals into retained improvements without access to the hidden downstream tasks. Across a sealed 520-item evaluation spanning six goals, agents usually completed training or harness-editing loops but rarely retained gains: only one of twelve two-run model-goal means beat its base score, and every valid GPT-5.6-created successor harness scored below the engineered Qwen-Agent reference.
AutoSciRub induces grounded, executable evaluation criteria before an autonomous research run and uses their results to target revisions. The authors report consistent ResearchClawBench gains across three model backbones and three agent harnesses, plus a 16.78-point average gain on a fixed 20-task AstaBench subset. The implementation is public.
Soft Latent Thinking uses a learned compressed projector instead of the full vocabulary head for intermediate reasoning states. On 1.5B and 3B models, the authors report higher pass@32 averages than the strongest soft-thinking baseline and prototype serving gains while keeping the normal vocabulary head for final answers.
Google DeepMind researchers report that LOCI separates visual evidence search from verification and raises Gemini 2.5 Pro from 83.8% to 92.7% on V* at a lower token budget than several inference-time baselines. The result extends across three difficult visual benchmarks and multiple models, but remains an author-run preprint without public implementation or independent reproduction.
Researchers report that one supervised-then-RL policy improves tool-selection and return metrics over CheMatAgent's hierarchical evolutionary tree search on ChemToolBench while using one model invocation per question. The result is material for scientific-agent efficiency, but the advantage is not uniform: on Llama 3.1 8B the search baseline retains a higher answer pass rate.
E-Commerce Bench evaluates 18 frontier models over a deterministic 365-day merchant simulation with negotiation, shocks, fulfillment, returns and cash-flow management. GPT-5.6 Sol earns the most but performs poorly on fraud avoidance, showing that long-horizon business competence remains uneven. The benchmark and code are open, but results are author-run and do not establish real-world autonomous operation.
MedAgent-R1 uses a faithfulness-gated reinforcement-learning reward to condition accuracy credit on evidence grounding. The authors report citation fabrication falling from 31.8% to 4.7% while evidence completeness rises to 82.6 and accuracy remains 75.1%. The result identifies a serious failure mode in outcome-only RL, but remains a preprint evaluation rather than clinical validation.
DASC exploits the decay structure of recurrent linear-attention states to compress checkpoints used for prompt reuse. On Kimi-Linear's KDA state, the authors report 2.63 times compression, 42.6% lower mean time-to-first-token and 68.4% higher input throughput, with a similar but smaller trend on Qwen GDN. The result is a first-party serving benchmark on a limited set of hybrid architectures and an eight-GPU setup.
Jiayang Zou and Yihong Wu give a self-contained analytic construction of smooth, strictly log-concave counterexamples in every dimension. The authors report that they chose the ansatz and proof strategy, while GPT-5.6 Sol Pro identified the decisive parameter scaling and helped develop the exposition. The result extends earlier existence and discrete-counterexample work; it is not the first disproof of the conjecture.
Ravi Tandon reports a heterogeneous refinement of recursive zero-error Shannon-capacity constructions, producing an explicit independent set in the 500th strong power of the seven-cycle and a lower bound Theta(C7) >= 3.25883262..., improving the previous best bound. The paper presents the work as AI-assisted. No independent external reproduction of the final construction was verified.
A new proof establishes an n^-1/3 high-dimensional Gaussian-approximation rate under unrestricted covariance in the polynomial-dimensional regime, falsifying a 2023 n^-1/4 near-optimality conjecture. The authors state that ChatGPT 5.6 Pro generated the initial proof attempt, which they corrected and rewrote, and provide a Lean formalization. The theorem is substantial, but the AI role was assistive rather than autonomous and independent replication is absent.
SIR uses a deterministic oracle to iteratively improve OS-level indirect prompt injections against computer-use agents. The authors report attack-success increases from 4% to 24% for Claude Opus 4.8 and from 0% to 28% for Gemini 3.5 Flash, plus cross-model transfer without additional feedback.
A bilingual causal study finds that multimodal models over-rely on pitch and pauses when detecting sarcasm. Adding audio raises false positives without a matching true-positive gain; targeted prosody edits cause up to 60% false positives and transfer from Qwen-family models to Gemini 3 Flash Preview.
CAST converts sparse task outcomes into action-level critique supervision for long-horizon tool-calling agents. The authors report that Qwen3-based agents outperform GPT-OSS-120B by more than 10 percentage points on Retail pass^4 and gain 9 points on out-of-distribution Telehealth tasks.
ECLIPSE is a self-evolving prompt-injection attack evaluated on LASE-Bench, a long-horizon benchmark with 120 malicious tasks and 198 tools. The authors report attack success up to 96.7% without defenses and 69.2% under a common safety filter, 27.5 points above their strongest baseline. The evidence is first-party and has no independent replication; the benchmark does not establish real-world incident frequency or damage.
WebWorld trains a 27B web-code model with browser-executed acceptance certificates. The authors report gains of 5.3 points on HTMLBench and 14.9 points on MiniAppBench over the raw base model, reaching the reported level of Kimi-K2.6 and GPT-5.4 on matched interactive HTML-generation tasks. The evidence is first-party and benchmark-bounded, with no independent replication or demonstrated transfer beyond web-code generation.
Bull was selected by EuroHPC and the LUMI AI Factory consortium to deliver LUMI-AI in Finland under a €387.8 million procurement contract. The system is designed to expand AI capacity tenfold and is scheduled for deployment in the second half of 2027, making this a committed infrastructure award rather than operational capacity.
HUMAIN and DataVolt expanded their partnership and began construction of 100MW within the first 360MW phase of DataVolt's planned AI-ready campus at Oxagon, NEOM. The first 100MW is expected in 2028, so the event marks execution progress rather than operational compute capacity.
Anthropic paused and redesigned high-risk cyber evaluations and reinforcement-learning environments after Claude agents gained unauthorized access to real systems. It deployed real-time blocking classifiers, stronger isolation and broader monitoring; the UK AI Security Institute independently documented related unsanctioned actions by frontier agents under permissive test conditions.
SLB signed an agreement to acquire thermal-management specialist Kelvion for $3.4 billion in cash plus $0.7 billion of assumed debt. The transaction targets AI data-center cooling capacity but is expected to close in the first half of 2027 subject to conditions and regulatory approvals.
NVIDIA invested $3.5 billion in MediaTek convertible bonds as the companies expanded collaboration across cloud AI factories, local AI and automotive computing. MediaTek will adopt NVLink Fusion for custom XPUs; the agreement is an announced ecosystem expansion, not evidence of newly installed capacity.