PhysLoop Uses an LLM to Make Video Generators Follow Physics

·4 views

Researchers unveiled PhysLoop in November 2026, a system that pairs a large language model with a video generator to enforce real-world motion laws. Unlike post-hoc filters, it injects physics reasoning during synthesis and lets the generator correct itself in repeated loops. A differentiable physics engine scores each attempt and returns specific corrections, which the language model uses to revise its reasoning. In the described example, a rolling-ball and toppling-block scene improved from a consistency score of 0.41 to 0.89 after four iterations, running at 16 fps on one RTX 4090 GPU. Corrections are stored in a replay buffer so later training can reduce the number of loop iterations needed.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

ReA-Bench Introduced as Benchmark for Auditing Responsible AI Agents

A DEV Community post presents ReA-Bench, a benchmark for evaluating how responsibly AI agents can be audited, as reported by arXiv AI. The piece is a systems-focused write-up from ShVaSa Lab that discusses the work through the lens of infrastructure and hardware efficiency. It includes a sample Python pipeline runner and a flowchart outlining input preprocessing, a compute kernel, and verified output with telemetry. The authors report performance figures comparing an unoptimized baseline with their optimized architecture, citing reduced memory footprint, lower p99 latency, and improved symbolic invariant compliance. They also recommend that engineers weigh memory sizing, quantization trade-offs, and latency bounds when deploying such workloads.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Staged Retrieval and Deterministic Keys Prevent Duplicate Candidates in Search

The article argues that deduplication should be built into retrieval design for recruiting candidate search rather than handled as a cleanup step after a model responds. It recommends staged retrieval with explicit collections, bounded queries, and a stable source key derived from tenant, document ID, and content version. A retrieval result should carry identifiers, access-control metadata, source references, chunk text, and a score so reviewers can trace ranking changes. The author compares Elasticsearch, Pinecone, Weaviate, and a thin vector API called Infrai, noting each option's operational trade-offs. Because queues are at-least-once, the piece stresses consumer idempotency so that retries after worker timeouts update existing chunks instead of creating duplicate candidate hits.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Coding agents leave git worktrees piling up and consuming gigabytes of disk space

Tools like Claude Code, Cursor, and Codex run tasks in separate git worktrees so several agents can work simultaneously. Each worktree is a full checkout with its own dependencies and build output, so these copies accumulate unnoticed. One Claude Code user reportedly found 27 forgotten worktrees occupying 80 GB, and a Codex user with a Rails app had 61 worktrees taking roughly 100 GB. The article explains where each tool stores its worktrees and offers commands to list their sizes, check for uncommitted or ignored files, and confirm whether work has been merged before removal. It also recommends using git's standard removal without the force flag, and lowering each tool's retention settings to limit future buildup.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Auditing SPF and DMARC across many domains: why a 'pass' can still mislead

Checking email authentication for hundreds of domains is often scripted, but many scripts report misleading results. The guidance covers reading SPF and DMARC TXT records correctly, including joining records split into 255-character chunks. It flags domains publishing multiple SPF records as broken and explains that a +all mechanism effectively means no SPF protection. It also notes that DMARC p=none only collects reports while enforcement requires p=quarantine or p=reject, and that a low pct value weakens enforcement. The article concludes with a safe record set for parked domains that send no mail and a Node.js approach for sorting domains into enforcement buckets.

0 comments Read more at DEV Community