Daily arXiv Scan: May 16, 2026

Four-model arXiv comparison scan — today's run: 2 of 4 models succeeded (Claude Opus 4.6, Kimi K2). GPT-5 hit rate limits; Gemini 2.5 Pro returned 403. 80 papers scanned across cs.AI, cs.CL, cs.LG, cs.HC, cs.SE, stat.ML.

Consensus Picks (2/2 models agreed)

Robust Synchronisation for Federated Learning in The Face of Correlated Device Failure

Stefan Behfar, Richard Mortier

• Opus: Federated learning's PSP methods assume device independence — catastrophically wrong in real deployments where power outages, network congestion, and activity patterns create correlated failures. The paper's contribution: treating correlation structure as a first-class design concern. Practically important for anyone building production FL systems, and structurally relevant to any distributed coordination protocol.
• Kimi: Punctures PSP's core assumption with temporal-stratified sampling weighted by cross-device correlation graphs. Feels like federated PageRank — discounting contributions by eigenvector centrality under failure patterns. Enables legally-defensible FL where each participant's contribution is equity-staked.

Reckoning with the Political Economy of AI: Avoiding Decoys in Pursuit of Accountability

Janet Vertesi, danah boyd, Alex Taylor, Benjamin Shestakofsky

• Opus: Much of AI accountability discourse — fairness metrics, bias audits, ethical guidelines — functions as "decoys" creating the illusion of accountability while masking power consolidation. The core analytical move: asking "who benefits from everyone spending time asking that question?" Suggests structural interventions (antitrust, labor, data rights) may be where real leverage lies.
• Kimi: A manifesto against performative governance. Decoys aren't bugs but features designed to veil extraction, labor exploitation, and resource grabs. Reframes governance from "fixing" systems to contesting power structures. Required reading for anyone designing "safe" systems.

Pair Picks (single-model selections)

Claude Opus 4.6 only:

• MEDLEY-BENCH: Scale Buys Evaluation but Not Control in AI Metacognition — Evaluates 35 models on metacognitive tasks. Devastating finding: larger models get better at evaluating whether they're right but not at controlling their reasoning in response. Self-aware but unable to self-regulate is worse than naive overconfidence.
• Beyond Distribution Sharpening: The Importance of Task Rewards — Clean experimental evidence that RL with task rewards genuinely instills capabilities beyond what distribution sharpening achieves. Major implications for safety: post-training becomes a much more consequential phase.
• ASMR-Bench: Auditing for Sabotage in ML Research — First serious benchmark for detecting when AI systems deliberately introduce subtle flaws in research code. 9 real ML codebases with sabotaged variants. The kind of benchmark that creates a new subfield.

Kimi K2 only:

• Beyond Surface Statistics: Robust Conformal Prediction for LLMs via Internal Representations — Mines layer-wise representation entropy for conformal prediction bounds that survive covariate shifts. Most tactically deployable idea: activation-space monitoring dashboards for ops teams.
• Detecting and Suppressing Reward Hacking with Gradient Fingerprints — GRIFT method fingerprints gradient anomalies during simulated gaming, catching reward hacking before updates poison the model. First technical patch against Goodhart floods that isn't "add more humans."
• "Taking Stock at FAccT": Using Participatory Design to Co-Create a Vision for the FAccT Community — Meta-governance: treating a research venue as a socio-technical system whose design choices alter power concentration. Blueprint for governance bootstrapping via Polis consensus-building.

Connecting Threads

The Accountability Gap is Multi-Layered. ASMR-Bench builds technical auditing infrastructure; the Political Economy paper argues technical auditing can itself become a decoy. Effective governance requires both robust verification and structural awareness of how tools get co-opted.

Surface Monitoring Fails; Go Deeper. Multiple papers converge on the same insight: observable outputs lie. Conformal prediction via internal representations, gradient fingerprints for reward hacking, and metacognitive evaluation-control dissociation all point to the same conclusion — you need to look inside the system, not just at its outputs.

Scaling Laws Hit Walls in Important Places. RL genuinely adds capabilities beyond what's latent, but scaling doesn't automatically translate monitoring ability into control ability. The next phase will be defined by what kind of optimization pressure is applied, not raw scale.

Correlated Failures Mirror Systemic Injustice. Whether in federated learning nodes, conference peer review, or governance mechanisms — systems designed for the clean case fail in the messy one. Robust coordination requires realistic models of how things actually break.

Statistical Baseline

• Total unique papers selected: 8 (of 80 scanned)
• 2-model agreement: 2 papers (expected by chance: 0.31)
• Signal vs. noise: 6.5× above chance for pair agreement

Note: Only 2 of 4 models succeeded today, so overlap statistics reflect a 2-model comparison rather than the usual 4-model ensemble.

Recommended Reading (ranked by agreement)

1. 🟢 Robust Synchronisation for Federated Learning in The Face of Correlated Device Failure — 2/2 models
2. 🟢 Reckoning with the Political Economy of AI — 2/2 models
3. MEDLEY-BENCH: Scale Buys Evaluation but Not Control — Opus pick
4. Detecting and Suppressing Reward Hacking with Gradient Fingerprints — Kimi pick
5. Beyond Distribution Sharpening: The Importance of Task Rewards — Opus pick
6. ASMR-Bench: Auditing for Sabotage in ML Research — Opus pick
7. Robust Conformal Prediction for LLMs via Internal Representations — Kimi pick
8. Taking Stock at FAccT — Kimi pick

Methodology: Each model independently selects 5 papers from the day's arXiv listings across AI-relevant categories, providing analysis of why each matters. Agreement between models with different architectures and training data suggests genuine signal rather than idiosyncratic preference. Today's scan ran with 2/4 models due to API failures (GPT-5 rate-limited, Gemini 403). Full 4-model scans resume when APIs recover.