Daily AI Briefing — September 7, 2026

AI SAFETY & ALIGNMENT

“Moral Competence Before Moral Content: Why LLM Agents Lack the Prerequisites for Coherent Alignment” (arXiv:2609.05036, accepted to the Paris Journal of AI and Digital Ethics / PCAIDE 2026) argues that alignment cannot meaningfully apply to current LLM agents because their moral behavior is not coherent in a structural sense — independent of any contested normative target. The authors, from the Aithos Research Foundation, begin from value pluralism: if there is no single correct set of values, the shared prerequisite for alignment is that a system expresses a coherent policy — a mapping from situations to verdicts that stays invariant while a situation’s morally relevant features are preserved and changes when they change. They operationalize this as four structural conditions (verdict stability, monotonicity, decisiveness, and Pareto viability) that are evaluable from behavior alone, without reference to a moral standard or expert baseline — a “structural floor” for alignment rather than a normative target. Testing nine frontier models on three simulated agent deployments under a factorial design (five paraphrases × five escalation levels × three dominance conditions), they report two striking results: no model expresses a coherent policy across all three deployments, and surface-form perturbation alone produces verdict-rate swings of nearly 100 percentage points at a single escalation level — meaning pure paraphrasing can flip a model’s moral judgment from one end of the scale to the other. Crucially, success on one scenario did not predict competence on another. The paper’s conclusion is deliberately strong: if coherence is the precondition for alignment, then “LLM-based agents are not currently the kind of object to which alignment can meaningfully apply.” [arXiv:2609.05036](https://arxiv.org/abs/2609.05036)

The significance is methodological as well as philosophical. Rather than another benchmark of which values models endorse, this is a framework for measuring whether models can hold any stable values — a distinction that matters for agents making consequential decisions, where verdict instability under paraphrase becomes a concrete reliability failure rather than an abstract ethics concern. The four-condition battery is cheap to administer and requires no ground truth, which makes it a plausible addition to agent safety evaluations.

AI EVALUATION

WearableQA (arXiv:2609.05405), from Meta and KAIST, introduces a benchmark for health reasoning over real longitudinal wearable data — 4,084 ten-option multiple-choice questions constructed from the wearable time series, blood biomarkers, and demographics of 200 real users, each contributing up to 500 days of daily measurements. Unlike synthetic health benchmarks, WearableQA preserves authentic device noise and inter-individual variability, and its 16 question types are organized along two complementary axes: data versus health reasoning (computation over longitudinal measurements versus physiological interpretation) and single- versus cross-signal reasoning (one signal versus integration of multiple signals). Question construction uses a dual-grounding framework that combines literature-grounded physiological findings with statistically validated population-level patterns. Across 14 proprietary and open-source LLMs, performance spans 19.6% to 72.9% against a 10% chance baseline, with most models below 60% — the benchmark is far from solved, and notably open-weight models trail substantially. The chain-of-thought ablations are the most diagnostic finding: CoT prompting delivers large gains for mid-tier models (Claude-Opus-4.6 +19.6 points, Gemini-2.5-Pro +18.6, GPT-5.4 +17.7), yet only +1.6 points for the top-scoring Gemini-3.1-Pro — suggesting that strong health reasoning skill largely obviates the benefit of explicit reasoning scaffolding, while weaker models depend on it heavily. Data is public via Meta’s GitHub and Hugging Face. [arXiv:2609.05405](https://arxiv.org/abs/2609.05405)

“Beyond Aggregate Scores: Behavioral Correctness Assumptions for Assessing Reference-Based Automatic Evaluation Methods” (arXiv:2609.05289) proposes a complementary meta-evaluation framework that probes how an automatic NLG evaluator behaves under controlled conditions, rather than only how often it agrees with human judgments. The authors define a taxonomy of correctness-preserving and correctness-altering assumptions, operationalized as controlled response transformations that specify expected scoring behavior — e.g., reformulations that should preserve a score versus edits that should lower it. Applying the framework across lexical, character-level, semantic, LLM-based, and hybrid evaluators, they analyze assumption-level behavior, stability, sensitivity, repeat-run variability, and configuration sensitivity. The contribution is a shift from correlation-style meta-evaluation — which can hide systematic evaluator failures behind aggregate agreement — to a behavioral test suite that exposes when and why an evaluator breaks. This speaks directly to the growing reliance on reference-based metrics and LLM-as-a-judge pipelines in agent and RAG evaluation, where a single-number agreement statistic rarely predicts performance on the failure cases that matter. [arXiv:2609.05289](https://arxiv.org/abs/2609.05289)

GLOBAL & GEOPOLITICAL AI

KOPA-Bench and the EDGE synthesis recipe (arXiv:2609.05395, EMNLP 2026 Industry Track) quantify the gap between open-source LLM agents and the multi-step tool-calling demands of sovereign government AI — and demonstrate a data-synthesis path to closing it, using 145 real-world tasks over live Korean public APIs. The motivation is regulatory: data-sovereignty rules increasingly require public institutions to deploy open-source, on-premise LLM agents over government APIs, yet these models consistently stumble on two characteristic failure modes — entity code-lookup chains, where simple queries decompose into dependent lookups, and high-cardinality responses, where agents must reduce or fan out across multi-record outputs. Models either skip prerequisite lookups or answer prematurely from partial results. To address the training-data bottleneck, the authors build EDGE (Execution-grounded Dynamic Graph for tool-calling data synthesis): it constructs a graph of how each tool’s output can feed another tool’s input, keeps only the links that actually succeed when called against live APIs, and traverses these verified links to synthesize executable multi-step trajectories. Fine-tuning a 9B model via GRPO on the resulting dataset nearly matches an untuned 27B model from the same family, with improvements transferring to the general-purpose Berkeley Function Calling Leaderboard (BFCL). The paper is notable both as a rare public benchmark over live sovereign infrastructure and as evidence that execution-verified synthetic data — not scale alone — is the binding constraint for on-premise agent competence. [arXiv:2609.05395](https://arxiv.org/abs/2609.05395)

TECHNICAL TRENDS

QuantumEvo (arXiv:2609.05327) uses an LLM as a heuristic generator rather than a direct predictor in reversible quantum circuit synthesis, discovering a variable-ordering heuristic that improves the quantum cost of circuits synthesized from binary decision diagrams (BDDs). Reversible circuit synthesis translates Boolean functions into quantum gates, and BDD-based approaches remain sensitive to variable ordering — but existing heuristics optimize BDD size, an imperfect proxy for downstream quantum circuit cost (QCC). QuantumEvo instead searches over ordering heuristics initialized from multiple heuristic families, selecting candidates by measured QCC. The discovered heuristic, HGA-QE, modifies the sifting step inside a genetic algorithm to better align with quantum cost, achieving a 70.9% tie-or-win rate against the per-function best baseline and strict wins on 13.5% of functions — with a clearer relative advantage on benchmark suites drawn from domains outside the heuristic-discovery data, indicating genuine generalization rather than overfitting. The broader pattern is the interesting part: LLM-generated algorithms winning on axes (here, quantum resource cost) that no human-designed heuristic targeted, via evaluation-driven search over program space rather than direct prediction. [arXiv:2609.05327](https://arxiv.org/abs/2609.05327)