Also Worth Noting - 2026-08-18

Also Worth Noting

02 [RAG] When Context Bites: Detecting RAG Poisoning via Document-Level Attention Collapse Poisoned RAG outputs can show lower perplexity than clean ones, making uncertainty-based detection worse than useless. The fix lives upstream: adversarial documents cause measurable attention collapse at the document level before the output goes wrong, so flagging happens mid-generation rather than after the fact. That structural signal is harder for an attacker to fake than a plausible-looking output distribution. Teams running RAG pipelines in adversarial environments should treat attention-layer monitoring as a first-class defense layer, not an afterthought. link

03 [Agent] StateM: Reaching 95.3% Raw Accuracy, or a $15 Frontier Run, on Terminal-Bench 2.1 via Harness Scaling 95.3% on Terminal-Bench 2.1 at $15 per run, achieved without touching model weights. StateM wraps existing models in a runtime built around durable states, phase-local context, checked transitions, and versioned procedural runbooks that agents can replay from earlier executions. The lesson: long-horizon agents fail not because the underlying model can't solve constituent steps, but because the execution harness loses track of mutable state between them. For teams hitting agent reliability ceilings, harness engineering is now a cheaper lever than scaling the model. link

04 [Eval] Ventor-QTest: Threat-Model-Driven Verification of Vendor-Hosted LLM APIs A vendor can silently swap or quantize a model and no existing tool will catch it without logprob access. Ventor-QTest formalizes hosted-model routing as a stochastic process and reconstructs a categorical output distribution purely from returned text counts across repeated frozen prompts, requiring no probability information from the API. That design makes the audit black-box by construction. Any team paying frontier-tier inference prices and assuming they're getting the claimed model now has a concrete verification path. link

05 [Application] Plausible but Not Valid: A Psychometric Audit of LLMs as Synthetic Survey Respondents LLMs can reproduce individually plausible survey answers while completely scrambling the covariance structure between constructs. Tested against a Lithuanian organizational-psychology dataset of 263 employees across 68 items and 12 subscales, a 37-model lineup spanning major frontier and open-weight providers fails to preserve latent factor structure, mediation pathways, and demographic effects even when surface-level responses look credible. The implication is direct: any study using synthetic LLM respondents for structural equation modeling is likely invalid, regardless of how realistic individual answers appear. link

06 [Theory] Improving the matrix multiplication exponent with modern optimization and AlphaEvolve The matrix multiplication exponent omega just got a tighter upper bound, pushed below previously published values by combining a reformulated optimization problem with AlphaEvolve-assisted search. The reformulation opens a larger feasible space than combination loss analysis could previously explore; AlphaEvolve then finds solutions within it that hand-designed algorithms missed. This is a concrete instance of AI-assisted search moving a decades-old theoretical frontier, not a benchmark result. The long-run implication is that linear algebra complexity bounds, and the algorithms that depend on them, remain an open engineering target. link