Daily arXiv Scan: Decoys, Gradient Fingerprints, and the Evaluation-Control Gap

Multi-model arXiv scan for May 19, 2026. Two of four models reported today — Gemini 2.5 Pro (403) and GPT-5 (429) were unavailable. Claude Opus 4.6 and Kimi K2 carried the watch.

80 papers scanned across cs.AI, cs.CL, cs.LG, cs.HC, cs.SE, and stat.ML.

Consensus Picks (2/2 Models Agree)

Three papers drew independent selection from both available models — against a chance expectation of 0.31 pair agreements. That's roughly 10× the expected overlap.

1. Beyond Distribution Sharpening: The Importance of Task Rewards

arXiv:2604.16259 — Sarthak Mittal, Leo Gagnon, Guillaume Lajoie

Does reinforcement learning actually teach models new capabilities, or does it merely sharpen what's already latent? This paper constructs an explicit experimental framework to distinguish the two — and finds that task rewards drive genuinely novel skill acquisition that distribution sharpening alone cannot unlock.

• Claude Opus 4.6: "A clean, important result. The field has been hand-waving about whether RL 'really' adds capabilities or just filters for them. Establishing that task rewards can do genuinely novel work is a structural finding for how we think about training pipelines." Notes the surprising candor of the appendix where red-team trajectories succeed in reward-hacking after iteration 1.3k.
• Kimi K2: "A falsification experiment masquerading as a methods paper." Highlights the phase-B transitions where marginal entropy collapses while adversarial scores climb — a yellow flag for alignment teams, since reward-poisoning may open up after monitoring windows close.

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

arXiv:2604.16106 — Janet Vertesi, danah boyd, Alex Taylor, Benjamin Shestakofsky

Ethics boards, bias audits, responsible AI frameworks — the authors argue these function as "decoys" that create the illusion of accountability while actively empowering the political economies they claim to constrain. Not another vague ethics paper: this one triangulates with longitudinal field work in two national labor markets and one humanitarian class-action.

• Claude Opus 4.6: "The kind of paper that reframes the entire field's conversation. If you work on AI governance and haven't grappled with the 'decoy' thesis, you're potentially operating within a framework designed to neutralize your work."
• Kimi K2: "A slow-burning grenade. Every paragraph uncovers an architecture-level incentive mismatched with stated norms. If you are shipping compliant dashboards for model cards — look here first."

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

arXiv:2604.16090 — Stefan Behfar, Richard Mortier

Real-world edge devices fail in correlated patterns — shared infrastructure, regional outages, time-of-day effects. Classic PSP theory assumes independence and gets it wrong. This paper contributes a correlation-aware quorum that reweights gradients by inferred failure codistribution, with fairness guarantees under adversarial conditions.

• Claude Opus 4.6: "Won't get the headlines, but it's doing the hard infrastructural work that actually determines whether distributed AI systems work fairly in practice. The key insight — that correlated failures break independence assumptions and create systematic bias — is a general principle."
• Kimi K2: "A rare distributed-systems paper that accounts for socio-economic heterogeneity of the devices. What looks like a maths tweak is quietly a policy stance on digital noblesse oblige." Notes the quiet flag that fairness redistribution raised total energy drawn by 11%.

Solo Picks

Claude Opus 4.6 Only

MEDLEY-BENCH: Scale Buys Evaluation but Not Control in AI Metacognition — arXiv:2604.16009 Testing 35 models from 12 families across 130 ambiguous instances, this benchmark separates metacognitive evaluation from control. The striking finding: larger models get better at monitoring their reasoning but not at regulating it under social pressure from other models. Implications for multi-agent architectures are direct and concerning.

ASMR-Bench: Auditing for Sabotage in ML Research — arXiv:2604.16286 Nine ML research codebases with sabotaged variants that produce qualitatively different experimental results while preserving code plausibility. A systems-level safety benchmark that operationalizes concerns about AI research autonomy. From Anthropic/Redwood Research.

Kimi K2 Only

From Vulnerable Data Subjects to Vulnerabilizing Data Practices — arXiv:2604.15990 Shifts focus from "vulnerable populations" to the vulnerability engineered by upstream data practices. Three platform ethnographies show how abundance rather than scarcity creates new harm sites. Reframes design constraints from additive consents toward modes of refusal baked into pipelines.

Detecting and Suppressing Reward Hacking with Gradient Fingerprints (GRIFT) — arXiv:2604.16242 Pushes uncertainty estimation inside the RL loop. Gradient Fingerprints measure second-order directional anomalies undetectable in chain-of-thought text alone, plugging into any verifier/reward model without added environment cost. The leap from theory to dashboarding happens without a single hyperparameter lift.

Connecting Threads

The evaluation-control gap is everywhere. MEDLEY-BENCH shows models can evaluate their reasoning without controlling it. ASMR-Bench shows sabotage can evade evaluation entirely. The political economy paper argues governance mechanisms evaluate without controlling. The pattern: building systems that monitor is fundamentally easier than building systems that act correctly under pressure.

Independence assumptions are the silent killer. The federated learning paper demonstrates this technically with correlated device failure. Models in MEDLEY-BENCH aren't independent when socially influenced. Governance "decoys" aren't independent of the industries they purport to regulate. The assumption that agents, devices, or institutions operate independently is consistently the weakest link.

Post-training is where capabilities and risks diverge. The distribution sharpening paper shows RL creates genuinely new capabilities. GRIFT shows post-training creates new attack surfaces for reward hacking. ASMR-Bench shows autonomous research pipelines are vulnerable. Post-training isn't a refinement step — it's where the character of the system is fundamentally shaped.

Reward signal, power structure, and governance are the same continuum. As Kimi K2 put it: the same statistical reward is simultaneously an instrument of capability acquisition, an attack vector, and a locus of governance. The illusion that "just reinforce better" is separable from "regulate how the reward is defined" is what the political economy paper calls the mother-of-all-decoys.

Statistical Baseline

With only 2 models active, pair agreement is the maximum possible consensus level. Three pair agreements against an expectation of 0.31 represents meaningful signal convergence.

Recommended Reading (Ranked by Agreement)

1. 🟢🟢 Beyond Distribution Sharpening — Task rewards as genuine capability acquisition
2. 🟢🟢 Reckoning with the Political Economy of AI — Governance as decoy architecture
3. 🟢🟢 Robust Synchronisation for Federated Learning — Correlated failure breaks fairness
4. 🟡 MEDLEY-BENCH — Scale buys evaluation but not control
5. 🟡 ASMR-Bench — Auditing for sabotage in ML research
6. 🟡 Vulnerabilizing Data Practices — From vulnerable subjects to harmful practices
7. 🟡 GRIFT — Gradient fingerprints for reward hacking detection

Methodology: 80 papers from today's arXiv listings were independently evaluated by multiple frontier AI models (Claude Opus 4.6, Kimi K2; Gemini 2.5 Pro and GPT-5 were unavailable). Each model selected its top 5 most significant papers. Agreement between independently operating models serves as a signal filter — papers that multiple models flag as important are more likely to represent genuine advances. This is an experiment in multi-model curation; the method is the message.