The RAG frontier, 2024–2026

Evidence cutoff: 2026-08-09. This review uses final official proceedings when available and labels preprints and industry reports. “SOTA” always means a named task, corpus, model, and metric—not a universal architecture.

Executive result: the frontier is a policy and systems frontier

The fixed pipeline

query → one top-k search → concatenate chunks → one generation

remains a useful baseline, not the frontier. Current work changes the decision problem:

whether external evidence is needed;
when to retrieve during a reasoning trajectory;
what query, modality, index, or tool to use;
which evidence has downstream utility rather than surface similarity;
how to represent relationships, hierarchy, images, tables, time, and conversational memory;
when to stop, abstain, escalate, or route to long context;
how to attribute, update, delete, secure, and economically operate the resulting system.

A useful modern objective is therefore constrained utility rather than answer accuracy alone:

U(\pi)=Q(\pi) -\lambda_c C(\pi) -\lambda_l L(\pi) -\lambda_r R(\pi),

where retrieval policy $\pi$ controls tools and evidence, $Q$ is task/grounding quality, $C$ cost, $L$ latency, and $R$ security, freshness, privacy, or compliance risk. Different products choose different weights and hard constraints.

1. 2024: correction, routing, hierarchy, graphs, visual pages, and evidence utility

Corrective RAG (CRAG): grade retrieval before trusting it

Status: 2024 preprint, widely used architecture; do not confuse it with Meta's separate CRAG benchmark.

CRAG fine-tunes a T5-large retrieval evaluator. A high-confidence result is decomposed into smaller “knowledge strips,” graded, and filtered. A low score triggers query rewriting and web search. An ambiguous score combines local and web evidence. The generator is modular and can be Self-RAG.

Original v1 tables reported PopQA improving 40.3 → 59.3 and PubHealth 39.0 → 75.6 for a Self-RAG/Llama-2-7B setting; Self-CRAG reported PopQA 61.8 and biography FactScore 86.2. These values depend on evaluator thresholds, model, paper version, and web setup.

Why it matters. Retrieval confidence becomes an explicit control signal, not an assumption. Why it is not a general solution. The maximum document score is a brittle route criterion; evaluators drift across domains; web search is mutable, costly, and privacy-sensitive; filtered snippets can discard necessary context; generator-only latency omits search overhead. Original preprint.

RAPTOR: retrieve from a recursive abstraction tree

Status: ICLR 2024 peer-reviewed.

RAPTOR preserves sentence boundaries at leaf chunks, embeds them, reduces dimension with UMAP, soft-clusters with Gaussian mixtures, asks an LLM to summarize each cluster, re-embeds summaries, and recursively repeats. Retrieval either traverses the tree or searches a collapsed set of nodes from every level.

The paper's controlled NarrativeQA tree results improved ROUGE-L 29.26 → 30.87 and METEOR 18.15 → 19.20. On QASPER, GPT-4 with RAPTOR reported F1 55.7 versus DPR 53.0 and BM25 50.2. Its best configuration retrieved about 2,000 tokens.

RAPTOR is compelling when the answer combines themes at multiple scales. It is less attractive for high-churn corpora: insertions and corrections propagate upward; clustering is stochastic; summaries can omit or invent facts; index cost is shifted to ingestion; comparisons sometimes change reader and retriever together. ICLR paper.

Adaptive-RAG: classify complexity before choosing a retrieval budget

Status: NAACL 2024 long paper.

A T5-large classifier routes each question to no retrieval, one retrieval, or a multi-step IRCoT pipeline. Silver labels select the simplest strategy that answers correctly; if all fail, dataset-level assumptions provide fallback labels. Cross entropy trains the router.

With FLAN-T5-XL, HotpotQA Adaptive-RAG reported EM/F1 42.0/53.82 at 3.55 steps and 5.99 seconds versus always-multi 44.6/56.54 at 5.53 steps and 9.38 seconds. It approximately tied always-multi on MuSiQue but lost about nine EM on 2Wiki. The router confused “no retrieval” with one-step 47% and multi-step 22% in the reported analysis.

This is an accuracy/latency Pareto result, not the most accurate setting. Labels are noisy and partly encode dataset identity; the three-action menu cannot discover new strategies; distribution shift changes complexity. Its enduring idea is learned budget routing. NAACL paper.

Microsoft GraphRAG: global corpus sensemaking

Status: Microsoft Research industry preprint/official implementation, first public April 2024; not a general fact-QA leaderboard result.

Index construction uses an LLM to extract entities, relationships, claims, and descriptions. A weighted entity graph is partitioned with hierarchical Leiden communities. The LLM writes bottom-up community reports. For a global query, reports are shuffled and batched; parallel map calls produce partial answers and helpfulness scores; low-score outputs are removed, ranked, and reduced.

On 125 generated global-sensemaking questions for each of two corpora, GraphRAG variants beat vector retrieval for LLM-judged comprehensiveness in roughly 72–83% of pairwise comparisons. Root-level community reports used over 97% fewer context tokens than source-text map-reduce; vector search was often more direct. Generated questions, only two corpora, and GPT-4 judgments make the scope narrow.

Dynamic community selection later reported about 77% cost reduction at one hierarchy level with similar quality, but some comparisons had only 29 queries. DRIFT search seeds local refinement from a global community view.

Use GraphRAG for global themes, corpus-wide trends, and relational synthesis. Do not assume it improves local facts. Entity resolution, LLM extraction, community stability, expensive indexing, incremental updates, and deletion propagation are first-order system problems. Original report, dynamic selection, and DRIFT.

HippoRAG: associative graph memory with Personalized PageRank

Status: NeurIPS 2024 peer-reviewed.

HippoRAG runs OpenIE over passages, builds an entity/relation graph, links query entities as seeds, applies Personalized PageRank, and maps high-scoring graph nodes back to passages. It reported up to 20% gains over prior RAG methods on multi-hop QA and 10–20× lower cost/6–13× faster search than IRCoT in the paper's settings.

Unlike GraphRAG's community reports, HippoRAG targets multi-hop associative retrieval. OpenIE and entity-linking errors can sever paths or join unrelated entities; graph construction is costly; PageRank relevance is not proof of logical support. NeurIPS paper.

RankRAG: one instruction-tuned LLM reranks and answers

Status: NeurIPS 2024 peer-reviewed.

RankRAG first instruction-tunes Llama 3 on roughly 128k general examples, then blends QA, reading comprehension, retrieval, and ranking data. At inference the same LLM reranks candidate contexts and generates. 8B/70B variants beat same-size ChatQA-1.5 on nine knowledge-intensive benchmarks and were competitive with GPT-4 on five biomedical tasks without biomedical fine-tuning.

Unifying reranking and generation can align evidence utility and amortize model management. It remains bounded by first-stage recall, uses substantial model compute for ranking, and can couple ranking errors with answer errors. NeurIPS paper.

RAG versus long context: route, do not declare a winner

Status: EMNLP 2024 Industry Track peer-reviewed.

“Retrieval Augmented Generation or Long-Context LLMs?” evaluated nine LongBench/∞Bench datasets with top-five Contriever/Dragon retrieval and Gemini-1.5-Pro, GPT-4o, and GPT-3.5. When complete input fit, long context averaged 7.6, 13.1, and 3.6 points above fixed RAG for those models; above 16k, RAG beat GPT-3.5 long context.

Self-Route answers from RAG first and asks the model whether context is sufficient; insufficient cases use the full input. It used 38.6% of Gemini tokens at 2.2 points lower score, 61% of GPT-4o tokens at 0.2 lower, and improved GPT-3.5 by 1.7. PassKey exposed benchmark fragility: a keyword query gave RAG 80.34 versus long context 65.25, but paraphrasing collapsed RAG to 4.58 while long context held 69.32.

The conclusion is conditional routing, not RAG replacement. Closed APIs, prompts, contamination, model self-calibration, a fixed top-five baseline, and 2024 price/context assumptions limit durability. EMNLP paper.

ColPali: retrieve page images with late interaction

Status: first public 2024; ICLR 2025 peer-reviewed.

ColPali renders each page as an image. PaliGemma produces about 1,024 patch vectors per page, projected to 128 dimensions; query tokens and page patches interact with a ColBERT score:

\operatorname{LI}(q,d)=\sum_i\max_j\langle E(q_i),E(d_j)\rangle.

LoRA training on 118,695 pairs uses pairwise/in-batch contrastive loss with a hard negative. On the ten-task English/French ViDoRe suite, average nDCG@5 was 81.3 versus 67.0 for the strongest reported parsed-text+BGE/caption pipeline. Indexing was 0.39 seconds/page versus 7.22 for parsing and query latency about 30ms in the reported setup.

The trade-off is index size: about 257.5KB/page versus 8.6KB for BGE. Threefold pooling cut vectors by 66.7% while retaining 97.8% of performance. The original work evaluates page retrieval, not grounded answer generation; some queries are synthetic and training-domain overlap exists. ColPali is a frontier retrieval unit for layout, figures, and tables—not proof that OCR is always inferior. ICLR paper.

RaLMSpec: accelerate retrieval without changing model outputs

Status: ICML 2024 peer-reviewed.

RaLMSpec speculates future retrievals, batches verification, prefetches, and uses an optimal stride scheduler. It reported document-level speedups of 1.75–2.39× for exact dense retrieval, 1.04–1.39× for ANN, and 1.31–1.77× for sparse retrieval across three LMs/four QA datasets; kNN-LM reached up to 7.59× exact and 2.45× ANN. The result holds semantic behavior fixed and attacks systems overhead, though speedups are hardware/workload-specific. ICML paper.

2. 2025: outcome-trained search agents, reasoning-aware retrieval, managed memory, and citations

Search-R1, ReSearch, and StepSearch: reinforcement learning controls search

The agentic-RAG learning problem can be written as an MDP. State $s_t$ contains the question, reasoning trace, and observations; an action $a_t$ may emit reasoning, a search query, or an answer; retrieval changes the state; the policy maximizes expected return:

J(\pi)=\mathbb E_{\tau\sim\pi} \left[R_{\text{answer}}(\tau) +\sum_t r_{\text{process}}(s_t,a_t) -\lambda\operatorname{cost}(a_t)\right].

Search-R1 (COLM 2025, peer-reviewed) applies outcome RL to an interleaved search/reasoning LM and masks retrieved tokens from the policy loss. Qwen2.5-7B and 3B improved 41% and 20% relative to same-setting RAG baselines over seven QA datasets. Outcome reward does not prove faithful reasoning or evidence use. COLM paper.

ReSearch (NeurIPS 2025, peer-reviewed) trains GRPO from scratch with <think>, <search>, and <result> and reports 8.9–22.4 absolute gains over listed baselines. Again, answer-only success can reward spurious trajectories. NeurIPS paper.

StepSearch (EMNLP 2025, peer-reviewed) adds information-gain and redundancy process rewards to stepwise PPO over 19k trajectories. It reported 11.2 and 4.2 absolute gains for 3B and 7B over search-RL baselines. Synthetic subquestions and proxy rewards remain gameable. EMNLP paper.

R1-Searcher remains a preprint. Its two-stage retrieval-tag then answer RL reports base-Qwen average CEM 60.6 versus SFT 50.1, but analysis finds reward hacking through fabricated document tags, no-search behavior, and malformed format. It is evidence for staged curricula and for the danger of proxy reward, not a settled peer-reviewed result. Preprint.

ReasonIR: train retrieval for reasoning utility, not topical similarity

Status: COLM 2025 peer-reviewed.

ReasonIR synthesizes triples containing a query, a genuinely helpful document, and a plausible but unhelpful hard negative. An 8B bi-encoder learns which text supports the required inference. It reported BRIGHT nDCG@10 29.9; explicit query reasoning plus an LLM tie-break reached 36.9. Downstream relative gains were 6.4% on MMLU and 22.6% on GPQA over closed-book settings.

The contribution is a target shift from semantic relevance to downstream reasoning utility. Embedding compute is high; synthetic teacher/domain bias can teach its own notion of helpfulness; retrieval metrics still need end-to-end validation. COLM paper.

GritLM: one model for embeddings and generation

Status: ICLR 2025 peer-reviewed.

GritLM unifies text embedding and generation in one model. The paper reports no task loss relative to separate specialists in its setting and over 60% RAG speedup for long documents by sharing model work. This is a deployment frontier: co-design representation and generation to reduce memory, loading, and repeated encoding. Unified optimization can also create interference and version-coupled indexes, so task/corpus validation remains necessary. ICLR paper.

HippoRAG 2: passages join the graph

Status: ICML 2025 peer-reviewed.

HippoRAG 2 adds passage nodes and contextual edges, integrates the query online, and uses LLM filtering. The paper's average QA F1 was 78.2 versus NV-Embed-v2 73.4, RAPTOR 65.6, and a reproduced HippoRAG 63.8; the abstract reports a 7% associative-memory gain. It uses Llama-3.3-70B for construction/reading plus NV-Embed-v2, so the graph contribution is not an inexpensive apples-to-apples comparison. ICML paper.

Long-context LLMs meet RAG: top-k has an optimum

Status: ICLR 2025 peer-reviewed.

This study finds performance can rise and then fall as $k$ increases because later candidates are harder negatives. Context reordering and RAG-specific fine-tuning mitigate the problem. Optimizing only recall@ $k$ is therefore wrong: evidence selection must include downstream distraction utility. ICLR paper.

Sufficient Context: combine answerability with self-confidence

Status: ICLR 2025 peer-reviewed.

An autorater labels whether supplied context is sufficient to answer. Combining that judgment with model self-confidence improves selective accuracy by up to roughly ten points at some coverage levels. This separates “I know the answer” from “the supplied evidence supports the answer,” a critical distinction for abstention. Autorater and confidence calibration errors still require labeled domain evaluation. ICLR paper.

Think&Cite: search while writing for claim-level attribution

Status: ACL 2025 long paper.

Think&Cite runs self-guided Monte Carlo tree search over query, retrieve, and write actions. Reflection guides expansion; generation-progress DPO and an NLI attribution reward score trajectories. GPT-4o on ASQA reported EM 50.1, citation recall 89.5, and precision 87.1 versus reranked RAG 42.1/83.4/82.3. Removing tree search yielded 42.1/78.2/75.0.

The result demonstrates that attribution can guide search, but roughly 30 iterations with three children plus external reward models is expensive. NLI and LLM citation metrics can be gamed and do not prove authority or provenance. ACL paper.

Managed conversational memory: RMM, M+, and ComRAG

RMM (ACL 2025) uses prospective reflection to summarize utterance, turn, and session at multiple granularities, then retrospective online RL with cited evidence to refine retrieval. It reported over 10% accuracy gain against no memory management on LongMemEval. Summary drift, privacy/deletion, and citation-derived reward bias remain. ACL paper.

M+ (ICML 2025) combines latent MemoryLLM memory with a co-trained retriever, extending reported retention from below 20k to above 160k tokens at similar GPU memory. Latent memory is less directly auditable, correctable, and deletable than an external store. ICML paper.

ComRAG (ACL Industry 2025) consolidates a historical QA store with dynamic centroids. Across three industrial CQA datasets it reported up to 25.9% vector similarity improvement, 8.7–23.3% latency reduction, and chunk-growth reduction from 20.23% to 2.06%. It is domain-specific evidence that write/consolidation policy matters as much as retrieval. ACL Industry paper.

Visual document RAG moves beyond retrieval-only evaluation

VisRAG (ICLR 2025) uses a VLM dual encoder over page images and a VLM generator, reporting 25–39% end-to-end gains over text RAG. Parsing baselines may be disadvantaged on layout-heavy tasks; visual-token cost and citation locality remain. Paper.
VDocRAG (CVPR 2025) adds self-supervised visual retrieval and dynamic compression of dense visual tokens and introduces OpenDocVQA. Retrieval, generation, token compression, and page/span attribution need separate accounting. Paper.
MoLoRAG (EMNLP 2025) builds a page graph traversed with a lightweight VLM, reporting +9.68% QA accuracy over direct LVLM and +7.44 retrieval precision over listed baselines on four DocQA datasets. Paper.
REAL-MM-RAG (ACL 2025) supplies realistic multimodal queries/labels and rephrasing, exposing table and paraphrase weaknesses. Paper.

syftr: search the pipeline Pareto frontier

Status: AutoML/PMLR 2025 peer-reviewed.

syftr uses Bayesian optimization and early stopping over agentic and non-agentic pipeline configurations, jointly optimizing task quality and cost. Across its RAG benchmarks it reports configurations averaging 9× lower cost while preserving most accuracy of the most accurate Pareto-front flow. This is an important methodological direction: component choices and hyperparameters interact, so pipeline architecture itself can be optimized—provided the search objective contains risk, latency, and robust held-out data rather than one quality score. PMLR paper.

3. 2026: retrieval becomes a learned action inside reasoning

GRIP: retrieval as generation

Status: ACL 2026 long paper, July.

GRIP expresses retrieval decisions as output control tokens such as retrieval, intermediary reasoning, answer, and solved. Self-Triggered Information Planning decides when to retrieve, how to reformulate, and when to terminate inside one autoregressive trajectory. A Llama-3-8B model receives 40k supervised examples and 5k DAPO RL examples; reward combines BLEU answer fidelity and control-token accuracy.

With BM25 top three and at most three calls, mean score across HotpotQA/PopQA/NQ/WebQuestions/TriviaQA was 41.0 versus strongest listed open baseline InstrucRAG 37.0 and GPT-4o 41.4. GRIP without RL scored 40.7: most gain came from structured supervision/control rather than RL. Raising maximum calls three → ten increased mean calls 1.24 → 1.62 and score only 41.0 → 41.8.

This is strong evidence for unified retrieval control and diminishing search returns, but it uses short-form QA, static Wikipedia/BM25, and teacher-synthesized trajectories; it does not prove citations, security, or freshness. ACL paper.

Q-RAG: learn evidence selection while freezing the LLM

Status: ICLR 2026 oral.

Q-RAG treats evidence selection as value-based RL. State concatenates query and selected chunks; actions select another chunk or STOP; a shared embedder scores

Q(s,a)=\langle E(s),E(a)\rangle.

Sparse terminal reward measures whether all supporting facts were collected. The generator stays frozen. Official results claim named-task SOTA on BabiLong and RULER to 10M tokens and competitive HotpotQA/MuSiQue. The repository says a run fits one A100 80GB in about 12 hours but also warns active refactoring.

Q-RAG isolates retrieval-policy learning and scales to huge candidate context. It depends on support-fact supervision and synthetic/long-context tasks; it does not directly optimize citation truth or real-world datastore operations. ICLR page and official repository.

DeepRAG and HiPRAG: learn search decisions and process quality

DeepRAG (ICLR 2026) casts decomposition plus retrieve/reason choices as an MDP; its official abstract reports +25.41% answer accuracy with greater retrieval efficiency. That percentage is relative to paper-specific baselines and must not be compared across papers. ICLR page.

HiPRAG (ICLR 2026) adds hierarchical process reward: the fraction of optimal search/non-search steps, on top of outcome and format rewards. Across Qwen2.5/Llama3.2 and seven QA benchmarks, mean accuracy was 65.4% for 3B and 67.2% for 7B; over-search fell from above 27% in prior baselines to 2.3% while under-search also decreased. It is strong evidence for process supervision but depends on defining and labeling an “optimal” action. ICLR page.

Knowledgeable-R1: learn when retrieved text conflicts with parametric knowledge

Status: ICLR 2026 peer-reviewed.

Knowledgeable-R1 samples paired trajectories with and without retrieval, uses local and global advantages, and applies an asymmetric transformation to learn when to trust parametric knowledge over misleading context. The official report states +22.89% against baselines in counterfactual-conflict settings with no loss on fully accurate context.

This attacks context conflict and over-trust, not source provenance or indirect prompt injection. Parametric knowledge can itself be stale or wrong; calibrated conflict resolution needs authoritative/time-aware sources. ICLR page.

LDAR: select evidence for downstream utility under distraction

Status: ICLR 2026 peer-reviewed.

LDAR learns distraction-aware retrieval across six knowledge-intensive benchmarks. Its central finding is that top- $k$ coverage is not monotonic utility: adding a topically relevant hard negative can hurt the reader. This formalizes the lesson from Lost in the Middle and 2025 long-context work. A retriever must be evaluated jointly with the reader and evidence budget, not only against qrels. ICLR page.

RAS and GraphRAG-Bench: build graphs selectively

RAS (ICLR 2026) interleaves targeted retrieval with a query-specific, incrementally constructed knowledge graph. It reports up to +8.7% for proprietary and +7.0% for open models on seven knowledge-intensive benchmarks. It avoids always building a global graph but pays per-query extraction and reasoning cost. ICLR page.

GraphRAG-Bench / “When to use Graphs in RAG” (ICLR 2026) evaluates graph construction, retrieval, and generation across fact retrieval, complex reasoning, summarization, and creative generation and finds GraphRAG frequently loses to vanilla RAG. The proper conclusion is conditional: relational, hierarchical, or global tasks can benefit if graph construction is accurate; flat fact lookup often does not. ICLR page.

RouteRAG: learn to route between text and graph evidence

Status: Findings of ACL 2026.

RouteRAG applies reinforcement learning to choose and combine unstructured text and structured graph retrieval as reasoning unfolds, replacing fixed hybrid pipelines. This is a natural extension of adaptive RAG from budget routing to representation routing. It inherits graph-construction quality, policy-transfer, and cost risks. ACL paper.

PROGRAM and PropRAG: structured paths without a monolithic graph policy

PROGRAM (Findings ACL 2026) generates structured retrieval programs and augmented multi-queries for multi-hop problems rather than relying on surface similarity. Paper.

PropRAG (EMNLP 2025) stores context-rich propositions instead of collapsed triples and performs LLM-free beam search over proposition paths. It reports zero-shot Recall@5 and answer F1 SOTA within its 2Wiki/HotpotQA/MuSiQue setting. Propositions preserve context but add extraction/indexing cost and inherit source-segmentation errors. Paper.

MegaRAG and RobustVisRAG: structured and robust multimodal retrieval

MegaRAG (ACL 2026 long) constructs a multimodal knowledge graph with text, visual, and spatial cues and uses them in retrieval and generation. It reports gains on global and fine-grained text/multimodal QA against listed baselines. The cost and failure modes of extraction, visual grounding, entity resolution, and graph maintenance make it a task-specific frontier, not a generic default. ACL paper.

RobustVisRAG (CVPR 2026) separates distortion and semantic causal paths and introduces Distortion-VisRAG with seven domains, 12 synthetic and five real degradation types. It reports +7.35 retrieval, +6.35 generation, and +12.40 end-to-end on real degradation with comparable clean performance. This is a frontier for scans, camera images, compression, and damaged documents; exact robustness remains tied to covered distortions. CVPR paper.

CompactDS: datastore quality can dominate agent complexity

Status: ICLR 2026 peer-reviewed.

“Frustratingly Simple Retrieval Improves Challenging, Reasoning-Intensive Benchmarks” builds a broad, high-quality datastore with in-memory ANN and on-disk exact retrieval. Its minimal RAG reports relative improvements of 11%/34%/26%/14% on MMLU/MMLU-Pro/GPQA/MATH for 8B–70B models and subsecond single-node retrieval.

The result is a warning against architecture theater: data breadth, quality, deduplication, and efficient exact verification may matter more than an agent loop. Corpus licensing, contamination, and source authority must still be audited. ICLR page.

RAG-RL: train citation and reasoning under increasing distractors

Status: Findings of EACL 2026.

RAG-RL uses curriculum learning from fewer to more distractor documents and rule-based rewards so a multi-hop answer generator both answers and identifies relevant evidence. The paper reports gains in answer and citation accuracy on three open-domain multi-hop datasets and analyzes sample ordering/rewards. It shifts some precision burden from retriever to reader, but retriever recall remains a ceiling and rule rewards can be exploited. EACL paper.

4. A state-of-the-art taxonomy by problem—not by brand name

Problem	Frontier technique family	Strong baseline/control	Main risk
Exact fact/entity lookup	BM25/learned sparse + dense fusion + reranker	BM25 alone, dense alone	rare-term loss, stale source
Semantic zero-shot retrieval	strong bi-encoder, HyDE/query reasoning	BM25/hybrid	query drift, teacher bias
Reasoning-relevant evidence	ReasonIR, RankRAG, LDAR	relevance retriever + cross-encoder	hard negatives distract reader
Multi-hop QA	iterative/agentic retrieval, proposition or local graph paths	IRCoT / fixed multi-query	cost, error propagation, reward hacking
Global corpus themes	GraphRAG community reports, hierarchical summaries	map-reduce and vector RAG	expensive/stale lossy index
Long documents	RAPTOR, evidence utility selection, long-context router	full context and flat chunks	summary loss, positional distraction
Visual documents	ColPali retrieval; VisRAG/VDocRAG generation; robust variants	OCR+layout text hybrid	index size, visual token cost, attribution
Dynamic facts	time-aware live retrieval, snapshot replay, correction	static index	stale caches, non-reproducible web
Conversational memory	rewritten queries + multi-granular memory + consolidation	last-N turns, flat vector memory	privacy, deletion, summary drift
Citation-heavy answers	claim-level search/write and entailment checks	reranked RAG with spans	metric gaming, source authority
High-security corpus	provenance/ACL filtering + robust aggregation + isolation	trusted-corpus-only baseline	poisoning, injection, exfiltration
Cost-constrained product	learned routing/early stop, Pareto search, caching	fixed one-shot RAG	policy drift, hidden tail latency

No row has one universal winner. The correct experiment compares the technique with the strong baseline in the same row under the same corpus, generator, budget, and risk gates.

5. Open research problems

Faithful credit assignment. Outcome reward permits accidental success, fabricated evidence, and non-causal reasoning. Process reward imports its annotator's errors.
Policy transfer. Retrieve/stop/rewrite policies often learn one corpus, retriever, model, and benchmark distribution.
Claim-level provenance and authority. Citation syntax, NLI entailment, and source reliability are different axes.
Continual correction and deletion. Vector indexes, lexical indexes, graphs, summaries, caches, and learned memories must agree after updates.
Security under untrusted evidence. Poisoned content and indirect instructions can manipulate both retrieval and generation; tool-using agents amplify the impact.
Calibrated abstention under absence and conflict. Self-confidence alone is insufficient; sufficiency, authority, temporal validity, and contradiction all matter.
Realistic multilingual, multimodal, temporal, and multi-turn evaluation. Current tests remain small, synthetic, English-centric, or judged by LMs.
Joint heterogeneous retrieval at acceptable cost. Sparse, dense, visual, structured, and live sources have different indexes, scores, and SLAs.
Auditable memory. Consolidation and forgetting must preserve provenance, user controls, retention policy, and deletability.
Counterfactual component evaluation. Separate retrieval recall, evidence utility, reader robustness, generator parametric leakage, and citation use.
Cost, latency, memory, and energy reproducibility. Report p50/p95 and dollars per successful supported answer, not a model-only mean.
Evaluator validity. LLM judges, synthetic labels, and benchmark contamination can reverse system rankings.

The companion evaluation and risk review turns these problems into measurable gates; the decision guide maps them to an implementation workflow.