Each section below summarizes one paper's reported method, experiments, and limits from its full review. The studies use different tasks and metrics; their results are not directly comparable.
arXiv:2611.07393
CacheGraft: Enhancing Retrieval-Augmented Generation via Latent-Space Document Enrichment
CacheGraft repurposes cache-to-cache translation for retrieval-augmented generation by precomputing each document's fused key-value representation with a larger 4B sharer and a smaller 0.6B receiver. At query time, only the receiver loads the document cache and answers the question, avoiding execution of the sharer. Across AdversarialQA, SQuAD, CovidQA, TriviaQA, and QASPER, the authors report improved answer quality over receiver-only text RAG in 29 of 30 comparisons, with 17 statistically significant improvements. The approach is promising for small-model inference, but its storage cost, single-document retrieval design, and dependence on a specific C2C fuser remain substantial constraints.
How it works
The system represents a RAG prompt as x = [p,d,q], where p is a fixed preamble, d is the retrieved document, and q is the question. Offline, the sharer and receiver independently prefill [p,d]. A terminal layer-alignment strategy maps receiver layers to sharer layers, and a lightweight per-layer fuser combines their KV states. The resulting cache has the receiver's sequence length and representation space, so the sharer is not needed after indexing. The fusion operation is defined by Equation 1.
What the paper evaluates
The evaluation uses SQuAD v1.1, AdversarialQA, CovidQA, QASPER, and TriviaQA. The authors retain one question per gold passage, evaluate both dense top-1 retrieval and oracle retrieval, and run standard greedy decoding on one NVIDIA A100-40GB. The receiver is Qwen3-0.6B, the sharer is Qwen3-4B-Base, and the released C2C projector was never trained on retrieval data. Answer quality is measured with two locally run open-weight LLM judges, phi-4 and Gemma-2-9B, plus contains, which tests whether the generated answer includes the dataset gold answer. Higher values indicate better answer quality or answer inclusion.
Limit to keep in view
Storing a KV cache for every document can require orders of magnitude more space than raw text; the paper estimates roughly a terabyte for 100K documents without optimization.
arXiv:2611.00605
Conversational Interaction for LLM-Based Clinical Text-to-SQL: A User Study with Healthcare Professionals
This paper evaluates conversational versus single-turn LLM-based text-to-SQL for clinical EHR exploration with healthcare professionals. In a within-subjects study, 14 physicians used both interfaces over the MIMIC-IV Clinical Database Demo: mean SUS increased from 68.2 to 85.9 for the conversational interface, and participants consistently preferred it for satisfaction, adoption, and trust. Analysis of 183 dialogue sessions from 21 healthcare professionals showed that freely composed questions triggered clarification requests in 24.8% of cases, compared with 1.8% for predefined questions, with 64.3% of clarification episodes requiring EHR- or clinical-domain knowledge.
How it works
The main study used a counterbalanced within-subjects design in which 14 physicians used both interfaces. Phase 1 used three participant-selected questions from a predefined pool of 45 MIMIC-IV Demo queries: 15 requiring one table, 15 requiring two tables, and 15 requiring three tables. Phase 2 allowed unrestricted clinical or research queries.
What the paper evaluates
The evaluation used the MIMIC-IV Clinical Database Demo rather than a conventional broad text-to-SQL benchmark. The controlled phase used 45 predefined clinical queries spanning one-, two-, and three-table tasks; the open-ended phase captured naturally formulated questions. SUS scores range from 0 to 100, with higher values indicating greater usability. UTAUT measures covered Performance Expectancy, Effort Expectancy, Self-Efficacy, Anxiety, Behavioral Intention, and the average of the four positively oriented constructs. Paired t-tests or Wilcoxon signed-rank tests were selected after testing normality of paired differences.
Limit to keep in view
The main comparison involved only 14 physicians, so smaller subgroup effects and broader population differences may not be detectable.
arXiv:2611.03551
Can LLM Agents Be Forward-Deployed Engineers? Requirement Discovery and Delivery for Enterprise Text-to-SQL
This paper evaluates whether LLM agents can handle forward-deployed engineering for an enterprise text-to-SQL system: eliciting requirements from an ambiguous customer brief, then planning and executing post-training and deployment. In a case grounded in a completed Oracle-dialect engagement, the strongest discovery agents recovered 81.7% of withheld requirements, while only two of eight delivery runs met all three stated acceptance targets. Both passing runs still fell short of the human FDE's shipped system on the reported combined quality and production-validity results.
How it works
The reference engagement supplies annotated customer knowledge and deployment constraints. The discovery specification contains 17 core business, data, and evaluation requirements plus six decisions the customer had left open. A simulated customer, powered by Gemini 2.5 Pro, answers agent questions from a fact sheet; agents interact through ask, request-an-artifact, inspect, and submit actions under three ambiguity levels. Gemini 2.5 Flash judges whether withheld facts were disclosed and correctly included in the submitted requirements.
What the paper evaluates
For discovery, the authors evaluate 15 models across three ambiguity levels, with 10 independent runs per model and ambiguity level. Recovery rate is the fraction of withheld facts correctly submitted; elicitation count measures withheld facts both disclosed in dialogue and correctly submitted; ask turns count clarification actions, which may contain multiple questions. Claude Opus 5.5 reports the highest mean recovery at 81.7% ± 14.0, while Gemma 4 31B leads the open-weight group at 49.4% ± 17.2. A human audit of three models agrees with the LLM judge on 70.6% of recovery labels and 82.4% of disclosure labels.
Limit to keep in view
The study is a bounded case study of one enterprise Oracle-dialect text-to-SQL engagement; its findings may not generalize to other customers, domains, or communication styles.
arXiv:2610.40409
Retrieving from Your Own Mistakes: On-Policy Training of Embedding Models for Multi-Hop Retrieval-Augmented Generation
This paper shows that multi-hop dense retrievers fail partly because they are trained on clean, gold passage histories but tested on histories containing their own retrieval mistakes. Its on-policy state training procedure periodically rolls out the current retriever, labels every visited state with the earliest missing gold passage, masks other missing gold passages as non-penalized alternatives, and mixes these states with teacher-forced states in contrastive training. With the same 0.6B encoder and fixed document index, the method reaches 49.3 C@5 on MuSiQue, 20.8 C@5 on HotpotQA-Long, and 27.7 C@5 on 2Wiki-Long, while improving downstream MuSiQue exact match to 44.1 with a fixed reader.
How it works
A question q is paired with a corpus C and an ordered gold passage set G(q). The retriever serializes the question and previously retrieved passages into a state text, embeds that text with a query encoder, and compares it with document embeddings computed once by a fixed document encoder. Greedy retrieval appends the highest-scoring unseen passage at each hop. The supervision rule is defined for any state whose gold evidence is incomplete.
What the paper evaluates
The experiments train on 1,800 MuSiQue questions, select checkpoints on 200 questions, and evaluate on the official 1,000-question development set with an 11,656-passage corpus. The transfer benchmarks are HotpotQA-Long, with 100 test questions over 554,812 passages, and 2Wiki-Long, with 300 test questions over 220,000 passages. The retriever uses four chain steps on MuSiQue and eight on the long-chain benchmarks, with no further training on the transfer datasets. The evaluation reports recall and complete coverage at k = 5 and k = 10.
Limit to keep in view
HotpotQA-Long and 2Wiki-Long are constructed benchmarks, so transfer results may not represent naturally occurring long-chain retrieval distributions.
Read the full reviews
- CacheGraft: Enhancing Retrieval-Augmented Generation via Latent-Space Document Enrichment · arXiv:2611.07393
- Conversational Interaction for LLM-Based Clinical Text-to-SQL: A User Study with Healthcare Professionals · arXiv:2611.00605
- Can LLM Agents Be Forward-Deployed Engineers? Requirement Discovery and Delivery for Enterprise Text-to-SQL · arXiv:2611.03551
- Retrieving from Your Own Mistakes: On-Policy Training of Embedding Models for Multi-Hop Retrieval-Augmented Generation · arXiv:2610.40409
Source note: findings and limitations above come from the linked full paper reviews. Follow each original arXiv paper for the authors' complete methods and results.