SpatialCLI lets vision-language models drop external tools without major accuracy loss

·1 views

Researchers developed SpatialCLI, a training approach that enables vision-language models to internalize spatial reasoning capabilities previously dependent on external tool calls. The method boosted Qwen3-VL-8B-Instruct's accuracy on the MindCube navigation benchmark from 29.3% to 84.6% when tools were available, surpassing GPT-5.6 Sol's 72.1% score. After an internalization stage using reinforcement learning, the model retained 73.8% accuracy on MindCube even with all external spatial tools removed at inference time. The key practical benefit is eliminating per-call latencies of nearly three seconds, simplifying deployment infrastructure without sacrificing most of the tool-augmented performance gain. However, the authors note that evaluation remains limited to their own benchmark suite, and whether the internalized knowledge generalizes to broader environments like Habitat or RoboTHOR is still an open question.

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)

How to Run Local LLMs Privately Using Ollama and Spring AI

Developers can run large language models (LLMs) locally on their own machines using Ollama integrated with Spring AI and Spring Boot, eliminating the need to send data to cloud providers like OpenAI or Anthropic. This approach is particularly useful for building AI applications without API costs, working offline, and keeping sensitive company data within internal infrastructure. Ollama supports several open-source models including Llama, Qwen, Gemma, Mistral, and DeepSeek, and exposes a local API that Spring AI can communicate with directly. A practical use case demonstrated is an internal HR customer support assistant where employee queries are handled entirely on-premise. Developers are cautioned that running an LLM locally does not replace the need for proper authentication, authorization, and data security measures.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Details 5 Real-World Failures When Self-Hosting Chatwoot at Scale

A developer running self-hosted Chatwoot for around a dozen small Israeli businesses documented five production failures not covered in the official documentation. One major issue was that media attachments are stored in Docker volumes via ActiveStorage rather than in the PostgreSQL database, causing disk usage to spike undetected by standard database monitoring. The operator also found that duplicate media files — some stored up to 325 times — were consuming significant disk space, and a deduplication pass reclaimed over 3 GB across more than 7,000 blobs. A third failure involved the POST /messages API returning a 200 status immediately upon row insertion, before actual delivery to WhatsApp via Meta, causing a Sidekiq job queue to back up with over 3,700 jobs and nearly 11 minutes of latency during a resend loop. The account was handling several thousand conversations per week across two servers, with a custom drip-sequence engine contributing to several of the edge-case failures described.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Seven-Step Framework for Structuring IoT Test Automation Effectively

IoT testing is complex because connected products rely on hardware, firmware, networks, cloud services, and applications functioning in unison. A seven-step automation framework recommends starting by mapping device hardware revisions, firmware versions, and network conditions to define the full test environment. Teams are advised to prioritize stable, repeatable interfaces like APIs and protocol validation before automating more complex scenarios, while keeping some manual checks for physical setup. The framework emphasizes combining device simulators for scale with real hardware for physical validation, and integrating the entire test pipeline into CI/CD workflows so tests run automatically with each development change. Field failures should continuously feed back into the test suite, ensuring that resolved issues are covered by automation and are less likely to recur.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why Self-Correcting Agent Loops Are Essential for Reliable AI in Production

Large language models (LLMs) deployed in production environments are structurally prone to hallucination, meaning they generate plausible-sounding but factually incorrect outputs with no internal mechanism to flag uncertainty. This is not a prompt engineering flaw or pipeline bug — it is an inherent property of next-token prediction systems that conflate plausibility with truth. Researchers warn that fine-tuning and prompt optimization alone cannot eliminate this fundamental uncertainty, yet companies continue integrating LLMs into high-stakes customer, financial, and legal workflows. A proposed solution called the COPS framework — Self-Correcting Production Systems — treats the LLM as one component in a multi-step pipeline that includes verification, criticism, and re-generation loops rather than a standalone answer authority. Experts argue that meaningful human oversight in high-stakes decisions, combined with these self-correcting architectures, represents the only viable path to trustworthy production AI systems.

0 comments Read more at DEV Community