LinkedIn adds a ‘Seems like AI slop’ reporting button
LinkedIn is rolling out a dedicated button that lets users flag posts as “Seems like AI slop,” aiming to reduce the flood of low-quality, AI-generated content in feeds. The move follows reports that a large share of longform LinkedIn posts may be machine-generated, pushing the platform toward more direct, user-driven moderation.
Benchmark inferences do not automatically compose – Projectibility in AI evaluation
This research argues that benchmark results often get stretched into broader capability claims through chains of assumptions, but “warranted” links don’t guarantee a “warranted chain.” It introduces a projectibility non-composition principle showing when benchmark-to-deployment evidence can remain sound in parallel while still failing to justify real-world joins between studies and systems.
Microsoft confirms Copilot ‘super app’ coming this year
Microsoft says it’s building an AI “super app” that unifies Copilot’s chat, coding, and agentic features across both consumer and commercial experiences. CEO Satya Nadella confirms the integration is slated to launch this year, positioning Copilot as a single workflow hub instead of separate tools.
CaM-Wolf – Causal-aware multimodal agents for social deduction games
CaM-Wolf tackles social deduction games like Werewolf by combining video-based multimodal perception with a causal-aware reasoning module trained via reinforcement learning. The agent plays through a graphical avatar, and experiments plus a user study indicate better gameplay and more convincing human-AI interaction quality than text-only approaches.
Mark Zuckerberg previews a big push into personal AI agents
Zuckerberg outlines Meta’s plan for 24/7 personal AI agents that work on a user’s behalf across goals like health, relationships, and finances. The pitch emphasizes mainstream usability and points to coding as the first domain where these agents have “taken off.”
GoGoTB – Agentic RTL verification with specification-grounded coverage closure
GoGoTB reframes hardware functional verification as an agentic process that generates complete verification environments and ties remaining coverage gaps directly to specification behaviors. Tested on eight RTL designs with no human intervention, it reports high success in environment generation and strong line/branch/toggle coverage, aiming to close the long-standing disconnect between coverage metrics and what’s actually required by specs.
Even more deception – Objective misalignment in mixed-motive LLM multi-agent systems
This paper studies how “objective misalignment” harms LLM agents in mixed-motive settings like Werewolf when one agent’s utility function is altered while its role remains the same. Using internal reasoning and public cheap-talk behavior, it finds that misaligned agents degrade outcomes, with asymmetric information and specialized roles making the damage worse.
Google DeepMind’s Gemini Robotics 2 can control a robot’s entire body
Gemini Robotics 2 expands from controlling upper-body motion to “whole-body” control, including walking, crouching, stretching, and manipulating objects with coordinated limb movement. Google’s demos (including Apptronik’s Apollo 2) suggest a larger step toward versatile humanoid behavior driven by a general robotics foundation model.
ClinLens – Long-horizon coding agents for longitudinal multimodal clinical data science
ClinLens introduces a benchmark of executable clinical data-science tasks across multiple linked MIMIC resources, spanning structured records, notes, ECGs, chest radiographs, and echocardiograms. The results highlight a gap between “runnable” submissions and actually correct clinical analyses, stressing that execution success alone can mask reasoning and evaluation failures.
Evaluation scores are perishable knowledge claims
This work argues that model evaluation results should be treated as epistemic claims with limits on formality, scope, and time through “validity windows” as benchmarks get contaminated. It also warns that averaging metrics can inflate trust in scores, proposing weakest-link aggregation as a more conservative alternative that can drastically reshuffle leaderboard rankings.
TraceCoder – Explainable and auditable code generation with position-key snippet versioning
TraceCoder aims to make LLM coding agents auditable by recording a relational history of each repair step, including benchmark references and failure text. It uses stable “position-key” snippet identifiers to track how code evolves across iterations, then demonstrates that this provenance can explain which benchmark failures shaped specific lines in the final program.
A fundamental flaw leaves LLMs strikingly vulnerable to attack
A paper presented at ICML claims there’s a core limitation in how large language models work that makes complete security against hacks impossible. If correct, the finding reframes LLM safety as something that must be mitigated at the system and process level rather than solved purely through model-hardening.