arxiv arXiv cs.CL · 3d ago · src: 5d ago · research

Stale-document poisoning causes outdated retrieval to override correct model answers

from English

Researchers identify "stale-document poisoning," a temporal alignment failure in Retrieval-Augmented Generation (RAG) where outdated external evidence causes models to provide incorrect answers despite knowing the right response without retrieval.

The study constructs a benchmark of 317 verified knowledge reversals across medicine, law, software, and platform policy.
Outdated retrieval flips 30% of Llama and 37% of Qwen answers even without instructions to trust the document.
Explicit follow instructions raise poisoning rates to 66% for Llama and 75% for Qwen.
Poisoning ranges from 17-91% across four open models and domains, while up-to-date evidence is followed in 97-100% of trials.
A fixed recency-aware hybrid re-ranker reduces poisoning by 4.6-10.0 points when dates are accurate.

The findings indicate that reliable RAG requires selective trust, where models must determine not only what retrieved evidence says but whether it still applies.

Importance 2/3 arXiv cs.CL Evaluation & benchmarks Retrieval & RAG

Read original

media Hugging Face Forums · 40d ago · 26 views

Researcher releases Ghost Search, an MCP-native Tor search layer for RAG

A researcher has open-sourced Ghost Search, a privacy-preserving meta-search aggregator designed to reduce reliance on paid APIs for Retrieval-Augmented Generation workflows. The project combines 14 onion search engines with an MCP server interface, allowing direct integration into LLM agents without backend infrastructure or per-query fees.

media Hugging Face Forums · 41d ago · 19 views

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.

arxiv arXiv cs.CL · 44d ago · 24 views

Nemotron-3-Embed closes commercial tax gap in MuSiQue retrieval benchmarks

A new audit of multi-hop retrieval benchmarks reveals that NVIDIA's Nemotron-3-Embed-8B has closed the performance gap between commercially licensed embedders and the non-commercial NV-Embed-v2 anchor on the MuSiQue benchmark. Released on July 16, 2026, this model is the only commercially licensed, free-to-self-host option that performs indistinguishably from the academic standard.

media Hugging Face Forums · 70d ago · 17 views

Author proposes open k registry for evidence saturation in local LLMs

The author has published a pilot paper and an open registry on Hugging Face to measure the evidence-saturation point (k*) of language models. This metric determines the optimal number of evidence fragments to inject into prompts, as adding more context does not monotonically improve correctness and can increase cost or epistemic contamination.

arxiv arXiv cs.CL · 76d ago · 18 views

LLM-as-a-judge signals fail to optimize closed-loop table recognition

A study using FinTabNet and OmniDocBench demonstrates that LLM-as-a-judge signals are insufficient for optimizing closed-loop regeneration in table recognition. The research reveals that judge scores frequently tied, rankings were not reproducible, and the system failed to recover better candidates produced by iteration.