Literature Review Leads And Corrections

Literature Review: Leads, Corrections & Open Questions

Part 2 of the Literature Review: downgraded citations, rejected reads, and open questions.

Downgraded, do not cite without re verifying

"Coordination as an Architectural Layer for LLM Based Multi Agent Systems"

(arXiv:2605.03310). Originally read as evidence that coordination failures are architectural, not capability limited. In full it's a position paper on a 100 question pilot that didn't survive its own stated significance bar, non peer reviewed, non academic. The vocabulary is still reasonable, it was never evidence.

"Verified Multi Agent Orchestration" (VMAO)

(arXiv:2603.11445). Originally the most interesting lead in a batch for its DAG of sub question dependencies. In full: 25 queries, one model family for both execution and judging, no significance testing, code unreleased, and the paper's own text attributes most of its gains to a retry loop DeepDelve's completion checks already have, not the DAG structure. Downgraded to a name worth remembering, not evidence to act on.

Verified findings worth acting on

"Do Agents Need to Plan Step by Step?"

(Otani et al., Megagon Labs, ACM CAIS '26, peer reviewed). Single step planning (DeepDelve's own adaptive loop) shows no accuracy edge over full horizon planning, and full horizon sometimes wins while using 2 to 3x fewer tokens. Single step also gets stuck in repetitive identical tool calls far more often, 30 to 45 percent of instances versus 1.9 to 5.9 percent. This is the theoretical shape of DeepDelve's own repeated same problem loops. Candidate fix: force a full plan regeneration on repetition detection instead of a narrow nudge, not yet implemented. Caveat from the paper itself: tested on closed tool set QA, not open ended web research, where it says single step may still be better.

"Lost in the Middle"

(Liu et al., Stanford/Berkeley, TACL 2024, foundational). Models use information at the start or end of context well and significantly worse in the middle, and accuracy saturates long before retrieved document count does. A credible second cause, distinct from budget truncation, for DeepDelve's "content vanishes during synthesis" pattern: material placed mid context is hard to use even when it's within budget.

"PIVOT"

(Zhang et al., Amazon, arXiv:2605.11225, preprint). Plan/inspect/evolve/verify trajectory refinement; human in the loop feedback gets up to 94% relative improvement, autonomous variant keeps most of it at 3 to 5x better token efficiency than competing methods. More importantly, section 4.3 tested whether a bigger thinking budget fixes long horizon synthesis failures: it doesn't. Nearly all reasoning tokens fire on the first turn (planning); final answer generation gets almost none regardless of budget ceiling. Gives a third candidate root cause for content silently vanishing during synthesis: the model doesn't allocate reasoning there no matter how much is available. Ablations also show disabling the final verify step hurts more than disabling planning, consistent with DeepDelve's own emphasis on grounding checks over plan quality.

"Demystifying Reinforcement Learning in Agentic Reasoning"

(Yu et al., NUS/UIUC/Princeton, arXiv:2510.11701). Directly actionable for the GRPO recipe: real end to end trajectories beat stitched synthetic ones for SFT init; conservative clipping and strong KL penalties over constrain exploration; higher policy entropy, especially for small models, helps training efficiency; fewer, more deliberate tool calls beat frequent or verbose calling. One correction: the "4B beats 32B" headline doesn't hold, it compares a 4B model's tool augmented score against a 32B model's no-tool score in a different table, not a like for like number. The paper's own limitations section says it only tested 4B/7B models, no generalization claim beyond that class.

Read and rejected

"Silent Failure in LLM Agent Systems... The Entropy Principle"

(arXiv:2606.08162). Single author, no affiliation, no co authors. The "derivation" is prose dressed as math, the validation is circular curve fitting to its own data, and the paper reads as marketing for a proprietary product. Salvageable only as loose vocabulary (silent degradation, cross session drift), not as theory for DeepDelve's endgame collapse pattern.

Open questions, resolved

The endgame collapse mechanism question got a real answer via LoopGuard (arXiv:2604.10044, unreviewed preprint, several Chinese universities and Alibaba). Long context decoding can collapse into persistent repetition when a subset of attention heads locks onto a narrow suffix of history, which KV cache reuse then self reinforces; smaller models collapse earlier and more often, matching several of DeepDelve's own disqualified small candidates. Real limitation: LoopGuard only detects verbatim lexical repetition, not DeepDelve's actual worst pattern (task name churn, semantically repeated but lexically varied retries), and DeepDelve doesn't control the serving engine's KV cache anyway. So the mechanism is confirmed, but the fix has to live at the application layer, not adopted from this paper.

Two smaller items also resolved: an ATLAS style failure taxonomy is buildable from an existing trace folder rather than requiring live runtime hooks, a closer fit for DeepDelve's own run state history. A separate large scale coding agent misalignment study (16,000+ validated episodes) found two direct parallels to DeepDelve's own catalog (inaccurate self reporting matching "narrate instead of write," and constraint violation matching the exclusion enforcement bug class) plus a cross session persistence effect DeepDelve doesn't currently instrument.

Still open: verifying the AXPO reward mechanism against DeepDelve's actual writer prompt shapes before the next GRPO round, and checking whether disqualified bake off models show shorter or absent reasoning traces before failed tool calls, testable against existing logs, no new reading needed.

What's merged into the main repo's docs

As of 2026-07-20: the capacity floor number, both constraint tax findings and the routing classifier they motivate, the MAST taxonomy mapping, the three way "architecture amplifies, doesn't rescue, capability" convergence (capacity floor paper, PIVOT, ATLAS), the three candidate causes for content vanishing during synthesis, the comparative survey conclusion, and the GRPO training levers from "Demystifying RL." See References and ROADMAP.md. This page stays the working document for what isn't load bearing enough to merge yet, including the FLARE/reflection distinction and the AXPO reward shape caveat.