A real example: what you get
15 new papers close to the project “Grounding Answers in Retrieved Passages: A Study of Hallucinations in Retrieval-Augmented Language Models”, found in the announcements of its first days
- How close it is on each facet: question, method, data, results
- What it shares with your project and what stays your own
- What to do: cite it, set your project apart, contact the authors, hurry
- The prior art of the last three years, and the new papers of the teams you name
Partial overlap
PCG-RAG addresses the risk that partially supported entity hypotheses redirect retrieval toward the wrong evidence. This is closely related to the project's finding that retrieval loses its grounding benefit when passages concern a related but different entity.
PCG-RAG: Progressive Compositional Grounding for Retrieval-Augmented Generation in Complex Web Question Answering
What stays your own
The project measures unsupported answer sentences using NLI and compares retrieval with a generator without retrieval; PCG-RAG instead proposes a multi-stage framework and reports QA accuracy. Emphasize the controlled measurement of entity-mismatched retrieval and the separation of retrieval failures from generator failures.
The email you would have received · 15 new papers are close to your projects
The first results
-
PCG-RAG: Progressive Compositional Grounding for Retrieval-Augmented Generation in Complex Web Question Answering
Partial overlap
PCG-RAG addresses the risk that partially supported entity hypotheses redirect retrieval toward the wrong evidence. This is closely related to the project's finding that retrieval loses its grounding benefit when passages concern a related but different entity.
-
Search-by-Answer Bias in Retrieval Models
Partial overlap
This paper identifies a dense-retrieval failure in which familiar answer objects receive favorable rankings even in passages expressing an incorrect relation. It provides a related account of how retrieval can select misleading evidence rather than protect against unsupported answers.
-
Cited but Unused: Parametric Provenance in Retrieval-Augmented Generation
Partial overlap
Both investigate whether RAG answers are grounded in retrieved passages rather than generated from parametric knowledge. This paper directly examines NLI verification and finds that it cannot reliably distinguish actual context use from memory-based answers that agree with the context.
-
Faithful in Steps: Improving Generalization and Citation in RAG via Query Decomposition
Partial overlap
Both study grounding failures in retrieval-augmented question answering. QDRAG addresses implicit entity identification and distracting context, which are closely related to the project's finding that retrieval about the wrong entity eliminates grounding gains.
-
CLADE: Category-Level Answering from Deficient Evidence in Retrieval-Augmented Language Models
Partial overlap
Both investigate unsupported answers in retrieval-augmented systems when retrieved evidence is related to the question but does not establish the answer. CLADE shows that coarse-grained evidence can encourage incorrect guesses, overlapping with the project's finding that retrieval's grounding benefit disappears …
-
Reasoning over Non-Parametric Memory in Multilingual LLMs: A Cognitive-Based Analysis of Relational Knowledge
Partial overlap
Both investigate whether language models reliably use supplied evidence rather than relying on parametric knowledge. Rel-Bloom isolates evidence-conditioned reasoning and shows that providing facts does not ensure reliable use of them.
-
CURATOR: Scalable Fine-grained Evidence Selection for Retrieval-Augmented Generation using Decision Models
Partial overlap
CURATOR studies how selecting supporting evidence from retrieved passages affects downstream RAG performance. It is relevant to the project's investigation of retrieval quality and the generator's use of available evidence, although its abstract does not report unsupported-sentence rates.
-
Decision Models as Retrieval Gates Against Injection and Pollution
Partial overlap
Both projects investigate how unreliable retrieved material undermines retrieval-augmented answers. The pollution experiments show that filtering misleading sources can substantially improve answer accuracy, connecting retrieval quality to downstream reliability.
-
Where Did the Evidence Go? Diagnosing Proof-Set Loss in Long-Term Conversational Memory
Partial overlap
EviCover distinguishes failures to retrieve or retain evidence from failures to use it in answer generation. Its finding that improved evidence recall does not establish improved answer correctness is relevant to the project's distinction between retrieval quality and grounded generation.
-
ARMOR: Adaptive Curriculum Meta-Learning for Noise-Robust RAG Reasoning
Partial overlap
ARMOR studies how retrieval noise compromises RAG and separates robustness into recognizing knowledge boundaries, verifying retrieval, and reasoning despite noise. This connects to the project's investigation of misleading retrieved passages and the generator's use of evidence.
5 more results hidden
See a complete example, explained step by step →
