Today's AI search track is mostly about catching failures in the retrieval chain. RAGLog reads only run logs to find which RAG stage went wrong, EMBERS rewrites questions to find failing cases for Deep Research systems, and two more papers measure how a soft brand placement in web pages and poisoned documents in a corpus affect retrieval. On methods, QueryFlow updates the question representation in multi-hop retrieval as evidence is found, and a scaling-law study asks whether to add model size or data when labels are limited.
RAGLog reads only the API logs each run of a RAG pipeline leaves behind, with no reference answers and no code changes, and points out which stage of the run went wrong. This matters because a pipeline can go wrong in the middle and still give the right final answer.
The paper builds a multi-hop graph retriever, QueryFlow, that revises the question as it searches.
When labeled data is limited, how good a retriever can get follows a law you can compute ahead of time.
Quietly adding one sentence that mentions a brand to a web article can be enough to get that brand into a generative search engine's answer.
In composed image retrieval, having a multimodal LLM first represent the edited image in a few parts and then gather them into the retrieval vector works better.
Deep Research systems split a question into subquestions, search the web repeatedly and then write an answer, so a small change in wording can push an early slip all the way through to a wrong answer. The authors analyzed 1,932 execution traces from three representative systems and found nine recurring kinds of defects, spread over subquestion writing, online search and final reasoning. EMBERS starts from questions a system already answers correctly, rewrites the question or an intermediate search input to target these defects, and checks that the rewrite still asks for the same thing. With the same 24-hour testing budget, it finds failing cases 3.21 times as fast as the strongest baseline. Fine-tuning on these cases also makes the model's answers more consistent.
In financial QA, a model can pull the right number from the documents and still attach it to the wrong company, period, unit or accounting item. Letting the model check and freely revise its own draft often breaks content that was right. CAR first diagnoses without editing: it judges whether the draft omits something, binds a value to the wrong item, lacks support, miscalculates or misreads the question, and a regression model turns those judgments into the share of required points the draft misses. It repairs only when that share passes a cutoff, set where repair gains and losses cancel out on calibration questions, and it edits only the spot most likely to be wrong. Across three financial QA datasets and three open-weight models, it recovers 14.6% of the points first missed while breaking only 2.1% of the content that was correct.
Take one model trained only on images and another trained only on text. With no paired examples at all, can their embedding spaces be aligned? This paper finds that a coarse alignment is possible, because independently trained models often learn similarly shaped spaces. Without any pairs, finding one global rotation (an orthogonal map) that makes the two point clouds overlap as a whole is enough to roughly line up two unrelated sets of embeddings. With only a handful of pairs, the method clearly beats existing ones, and the alignment even supports text-to-image generation without paired data. For cross-modal retrieval, paired data may be less essential than assumed.
When a vision-language model is used for text-to-image search, the top images for a concept can systematically over-represent some unrelated attribute, and nobody knows in advance to look for it. SPELUNKA needs no predefined labeled groups and no fine-tuning. Given a concept, it captions related images and extracts entities and attributes from the captions. It then compares how often each attribute shows up in the model's top-ranked results versus the whole dataset, flagging clear over- or under-representation as a suspect association. On four datasets it recovers both injected and naturally occurring biases, and the attributes it finds predict which images land in a query's top 5% with an average AUROC of 0.93. Before shipping a retrieval model, it can surface biases nobody thought to test.
Outside text can only sway a RAG answer after the retriever lets it into the context. This paper stress-tests that step: it plants a few poisoned documents per query in the corpus and measures how much the ranking is disturbed. On SciFact and NFCorpus, just 5 such documents per query cut the nDCG of classical retrievers by 40 to 60 percent. Most of the damage comes from copying the whole query verbatim into the document, and the same holds for a dense model like BGE-small, where the poisoned documents all reach the top of the results. Filtering by word overlap blocks these documents but also clearly hurts normal retrieval. RAG defenses need to start at the retrieval step.