7B Searches the Web, BVB Recovers 53.7% of Facts

Featured

01 Training: The 7B Model Becomes a System

A 7B model cannot memorize the open web. ZGCM-1 does not try. It treats internal reasoning and external tool calls as substitutes for parameter memory.

To support a 256K context, it alternates sliding-window and global attention. Training expands the context from 16K to 64K, then 256K. The team co-optimizes architecture, training systems, and curriculum.

The paper says 16K pretraining reaches the same loss about 4.2 times faster. That measures time-to-loss efficiency, not inference speed or a 4.2-times gain in final capability. On math reasoning and agentic search, the authors report results competitive with some much larger frontier models. Tool-heavy tasks may narrow the parameter gap, but small models have not caught up everywhere. Small-model teams may gain more from the released stage weights, checkpoints, data recipes, training code, and experiment logs. Cross-task replication, effective use of 256K context, and tool-call costs still need independent testing.

Key takeaways:

• Treat external tools as part of a 7B model’s capability design, not an add-on.
• The 4.2-times figure measures 16K pretraining efficiency, not inference cost or final capability.
• Study the open training recipe and intermediate artifacts before adopting its long-context and agent-training choices.

Source: ZGCM-1: A Fully Open and Extremely Efficient Foundation Model for Math and Agentic Search

02 Safety: Guard What Agents Actually Do

Once computer-use agents enter browsers, terminals, and file systems, risk appears in runtime interactions. Prompt-and-response review can miss the action that causes harm.

HazardAuditor runs Claude Code, Codex, Hermes, and OpenClaw in controlled environments. It normalizes their different interaction logs into a common event representation. That lets one safety monitor train across frameworks. An evaluation platform can score one execution, but its verdict does not automatically become reusable training data for a deployable guard.

GuardPO converts deterministic safety outcomes into sequence-level advantages. It normalizes rationales and final verdicts separately, preventing long explanations from dominating the safety decision during gradient updates. The abstract reports gains of up to 16.5 percentage points over the strongest existing guard. Deployment tests still need to measure cross-framework transfer, real-time blocking costs, and false positives.

Key takeaways:

• Audit runtime interaction traces, not just prompts and responses.
• Distinguish evaluation platforms that score behavior from guards that can operate across frameworks.
• Measure generalization, latency, and false positives alongside the reported 16.5-point maximum gain.

Source: HazardAuditor: From Executable Threats to Safer Computer-Use Agents

03 Robotics: One Objective for Mind and Motion

PhysBrain 1.5 builds task-centered clips from human interaction videos. It pairs semantic and spatial context with recovered motion and later observations. Supervised fine-tuning mixes human demonstrations, robot trajectories, and simulated experience.

The model encodes language answers, end-effector motion, and dense visual targets as three types of discrete sequences. A shared autoregressive next-token objective trains understanding, action, and future-state prediction together.

The 8B model averages 72.5 across 28 embodied-understanding benchmarks. It sets open-source best results on 14 tests. Yet the abstract provides only qualitative examples for action generation and future-scene prediction. A shared representation may reduce model stitching and fragmented objectives, but the reported measurements do not prove that. Robot deployments still need quantitative tests of control accuracy, inference latency, and error accumulation over long sequences.

Key takeaways:

• Pairing semantics, spatial context, motion, and later observations is central to the pretraining design.
• A shared autoregressive objective does not prove that integration costs have fallen.
• Understanding benchmarks cannot replace measurements of control accuracy, latency, and long-horizon error accumulation.

Source: PhysBrain 1.5: From Vision-Language Models to Physical Foundation Models

04 Evaluation: Rebuild the Video, Don’t Just Answer

BVB replaces video questions with a harder deliverable: code for a runnable, renderable, inspectable Blender animation. The output must reproduce object geometry, motion trajectories, and temporal relationships.

That deliverable exposes a gap between perceptual similarity and factual reconstruction. The best model scores 88.6 on perceptual similarity but preserves only 53.7% of the video’s spatiotemporal facts. More reasoning improves visual similarity without closing the factual gap. Models can copy the appearance without building a consistent executable world model.

BVB also measures Blender programming skill and operates under cost limits. It should not serve as a pure measure of video understanding. Use it as a stress test alongside existing question-answering benchmarks.

Key takeaways:

• Executable reconstruction directly tests whether models preserve geometry, motion, and temporal relationships.
• High perceptual similarity does not guarantee accurate spatiotemporal facts.
• Separate video understanding from Blender programming skill and cost-related measurement bias.

Source: BVB: Benchmarking Agentic Video Understanding via Programmatic Reconstruction in Blender