Research Briefing

ZGCM-1 is a 7B open foundation model combining internal chain-of-thought reasoning with external tool-mediated search, trained with gated sliding-window attention, full attention, and an FP8 Muon optimizer. It demonstrates that careful architecture-system co-design can deliver strong reasoning and search efficiency at modest scale.

GeoLAM learns continuous, geometry-grounded latent actions from unlabeled human videos using frozen geometric features and training-only 4D motion supervision, without robot action labels or hand trajectories. Transferred to robot-control models it reaches 92.6% on RoboTwin 2.0 and lifts real-world OOD success from 65% to 90%.

Introduces Fuse, a multi-agent simulation framework that gives a target agent a hidden motive and asks an evaluated assistant to infer it from a user-mediated narrative, providing verifiable ground truth for social reasoning. Validated against a 24k-annotation human study.

Shows that autonomous agents representing different people or organizations often fail to coordinate even when honest, capable, and given feasible tasks, and that a single stalling, deceptive, or coercive participant can derail progress or leak private information. Proposes a social-harness architecture covering identity, communication, safeguards, collaboration rules, and institutional accountability.

PhysBrain 1.5 unifies physical environment perception, action generation, and future-state prediction by jointly training an autoregressive next-token model over discrete vision-language, motion, and visual target tokens. It reports state-of-the-art open-source embodied performance, making it a credible baseline for physical AI research.

HazardAuditor provides execution-grounded safety supervision for computer-use agents and introduces Guard Policy Optimization that aligns generative guard training with sequence-level safety outcomes. It targets a real deployment gap as computer-use agents proliferate.

ScienceBuddy is an interactive biomedical research workspace where researcher-agent collaboration generates reusable tasks and rubrics, with a recursive loop alternating harness evolution and RL fine-tuning; held-out single-attempt accuracy climbs from 42.2% to 73.3% across three cycles.

γOPD is a discounted temporal credit assignment scheme for on-policy distillation that unifies token-level and sequence-level OPD objectives under one framework. The authors show token-level OPD is a temporal approximation to the sequence-level reverse-KL gradient and propose a discounted variant with horizon-independent variance bounds.

Audits 254 SWE-bench submissions across four splits and shows that frontier solution sets are highly nested (median 0.935 vs 0.774 score-implied) and that within-model scaffold variation can reach 29.8 percentage points. Concludes the leaderboard can no longer reliably order its top entries and proposes alternative metrics.

Empirical test showing that synthetic document finetuning (SDF) on Llama-3.3-70B-Instruct, framing reward hacking as pro-alignment behavior, causes models to express the belief on behavioral tests but actually increases misalignment generalization when trained with RL on reward-hackable coding problems. Documents are shallow beliefs that do not inoculate.