media Hugging Face Forums · 2h ago · src: 1d ago · open_models

ThoughtDAG user removes one context edge to change LLM answer

from English

A user testing the ThoughtDAG interface found that removing a single edge from an LLM's context graph changed the model's output, even when the model had previously repeated a wrong claim.

In a controlled scenario, a source node initially showed 17% worse performance, which flipped to 17% better after a corrupted note was added.
With the contaminated note still wired in, the model repeated the incorrect recommendation until the edge carrying that branch was removed.
Removing the edge excluded the upstream node from the next request without deleting the earlier work, causing the recommendation to change.
ThoughtDAG is a local-first, MIT-licensed project supporting Ollama and OpenAI-compatible endpoints.

The author seeks criticism on experiment design and questions whether context control should remain human-directed or be delegated to the system.

Importance 1/3 Trust 1/3 Hugging Face Forums Local inference Retrieval & RAG

Read original

media Hugging Face Forums · 5d ago · 26 views

ThoughtDAG tests explicit graph context control for local LLMs

ThoughtDAG is an open-source, local-first workspace that treats conversation edges as the mechanism for controlling model context. Each question-and-answer exchange functions as a node, and only upstream wired nodes are serialized into the next model request.

media Hugging Face Forums · 3h ago · 1 view

Celestium Engine adds VRAM optimizations and CPU fallback for SDXL

The Celestium Engine has introduced several upgrades to improve SDXL performance on low-VRAM hardware. These changes focus on memory management, stability, and fallback mechanisms to ensure consistent generation even on constrained systems.

media Hugging Face Forums · 9h ago

SplatRAG 3+ achieves SciFact R@10 0.8804 with slang in memory

SplatRAG 3+ demonstrates that a local Gaussian-splat memory can maintain high retrieval accuracy on the SciFact benchmark while simultaneously storing 5,000 Urban Dictionary slang rows without degradation.

github llama.cpp · 12h ago · 1 view

llama.cpp b10541 adds --mmproj-device argument for multimodal projects

The llama.cpp project released build b10541, introducing the `--mmproj-device` command-line argument to allow users to specify the device backend for loading multimodal projection files.

github llama.cpp · 12h ago · 1 view

llama.cpp b10545 fixes Metal matmul out-of-bounds reads for K not multiple of 32

The llama.cpp project released build b10545, which addresses a critical bug in the Metal backend's tensor API matrix multiplication kernel. The update prevents undefined behavior and corrupted results when processing tensors where the K dimension is not a multiple of 32.