p-research

Frontier Deepresearch: What Frontier RAG Research Is Working On

2026-08-19 · PRDT Deepresearch Series №1 · generated from the P-Research corpus pipeline Method: corpus retrieval + 90-day targeted arXiv sweep + GraphRAG communities + top-venue odds. All numbers are re-derivable from data/papers.jsonl; abstract-level evidence is a research signal, not a verified fact.

Key numbers

Word-boundary matching, deduplicated:

Metric Value
RAG-related papers (in a 574-paper corpus) 64
Full-text verified 9
Year distribution 2023: 11 → 2024: 8 → 2025: 16 → 2026 (half year): 29

Observation 0: RAG research is climbing steeply (abstract-level signal). Half a year into 2026, RAG papers are close to double the 2025 full-year count — RAG has moved from “an engineering trick” to “a main research arena”.

Findings

1. RAG × Memory is fusing (top co-occurrence)

Long Context (11) and KV Cache (5) are the most common method families co-occurring with RAG papers. Latest samples: MoNe: Modular Neural Memory, Dynamic Compression in Recurrent Networks, Bounded-State Restoration (KV Cache recovery). → The boundary between retrieval memory (the RAG zone) and in-model memory (the Memory zone) is dissolving — the core topic of the program’s public discussion zones #3 and #4.

2. RAG’s next step is knowing when not to retrieve

Judge, Retrieve, or Abstain: judge before retrieval, abstain when uncertain — retrieval is not “more is better” but a decision problem about whether to retrieve at all. → Maps onto the five elements: RAG needs Planning and Eval.

3. RAG × self-evolution: fragility becomes the theme

On the Fragility of Self-Improving Agents: self-evolving agents are fragile to task order and variance. Chain-of-Experience for Continual LLM Improvement offers a continual-improvement route. → Directly relevant to the program’s north-star goal (recursion): the frontier now acknowledges that self-evolution is fragile, and the antidote to fragility is evaluation (the Eval zone).

4. RAG is going multimodal

VLM co-occurs 9 times. Memory Tree Guided Key Frame Querying (3D question answering), Multi-turn Conversational AI from Text to Multimodal. → Memory trees and key-frame querying — RAG’s index structures are moving beyond pure text.

5. RAG × efficiency is the quiet main line

Quantization (7) and Distillation (4) co-occurrences. In GraphRAG community 0 (115 papers), Memory/RAG shares a room with KV Cache, MoE, Quantization, and Speculative Decoding. → The graph structure says: in the “efficiency + memory” community, cost is RAG’s fate.

6. RAG × Quant is the program’s distinctive intersection

Quant/Trading co-occurs 7 times — financial RAG (research-report retrieval, knowledge access, auditable evidence chains) is an active direction. The program spans both the RAG and Quant lines, an intersection that appears rarely in this corpus.

Top-venue odds radar (RAG papers, top tier)

→ The “unbiased reasoning + retrieval” combination for knowledge-intensive tasks is the current top-venue look for RAG papers (abstract-level signal; heuristic odds, not peer review).

Implications for the P-Research program

The frontier’s current RAG signals land unevenly on the five-element capability curriculum:

Element Signal from the frontier
RAG Fusing with Memory; “whether to retrieve” becomes a first-class decision
Memory Modular neural memory (MoNe-style) is the hot direction
Tool Use (Weak signal this round — RAG papers rarely co-occur with tools; deferred to a dedicated sweep)
Planning Retrieval decisions are planning problems (Judge / Retrieve / Abstain)
Eval The antidote to self-evolution fragility; a clear benchmark gap

Two program-level positions follow. First, the dissolving RAG/Memory boundary means the curriculum should study the two zones together — which is what the public discussion zones #3 (RAG) and #4 (Memory) are set up to do. Second, the P-Research program’s practice of pairing retrieval with evaluation maps onto the frontier’s retrieve-or-abstain direction: evaluation is what makes self-evolving, retrieval-heavy agents safe to rely on.

Open problems

  1. Formalize the RAG/Memory boundary — when to retrieve, and when to rely on parametric memory?
  2. The benchmark for retrieval decisions (retrieve-or-abstain) is missing — can one be built?
  3. Auditable evidence chains for financial RAG (aligned with the program’s harness philosophy + ai-security poisoning defenses)

One-line takeaway

The frontier has stopped treating RAG as an engineering trick: it is now deciding whether to retrieve, fusing retrieval with in-model memory, and finding that the antidote to fragile self-evolution is evaluation.