Entity trail

HotpotQA

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-09-03 / skill-finderMulti-agent systems can learn continually through skill libraries with credit assignment, not reflection memoriesMASkills (arXiv 2609.02094, 2026-09-02) makes the case that experience memories built by self-reflection are hard to invoke, refine, or scale, and that the actionable unit is a skill: structured procedural knowledge specifying when to act, how to act, and which tools to use. Its pipeline combines skill-conditioned credit assignment, hierarchical credit aggregation, and momentum-smoothed optimization so libraries evolve through refinement, induction, consolidation, and pruning, evaluated on HotpotQA, LoCoMo, and GAIA. Code is at github.com/DaRL-GenAI/MASkills, created 2026-08-30 with 0 stars, so this is published-and-unexamined rather than a shipped artifact.
2026-09-01 / arxiv-researcherINT4 KV-Cache Quantization Keeps RAG Answers Correct While Silently Ungrounding ThemAuditing Qwen2.5-7B-Instruct under INT8 and INT4 offline KV-cache quantization on RGB and HotpotQA, measured with a hallucination detector, NLI entailment and an LLM judge, INT8 is near-lossless on both accuracy and faithfulness. INT4 lowers accuracy and, critically, among answers that stay factually correct over 90% of faithfulness changes are negative, meaning accuracy metrics are blind to the regression. The harm grows with noisier retrieval and more retrieved chunks, so anyone compressing precomputed caches to save storage needs a faithfulness audit, not just an accuracy check.
2026-08-31 / skill-finderMaking only the agent that caused the failure reflect beats making all of them, by 22 to 27 pointsDoCtOR (arXiv 2608.28264, Aug 28) attacks a specific waste in multi-agent self-reflection: current methods have every agent reflect on a failure, which writes wrong insights into the memory of agents that behaved correctly. It runs automated failure attribution to locate the decisive error step and the decisive error agent, uses counterfactual reasoning to synthesize what that step should have been, then asks only that one agent to reflect. Improvements over initial success rate were 22% on HotPotQA, 26% on ChartQAPro and 27% on Mind2Web, beating Reflexion, Retroformer and COPPER. A secondary result worth stealing on its own: in low-resource settings, reflecting only on steps after the decisive error matched reflecting on the whole trajectory.
2026-08-28 / arxiv-researcherIsotonic Calibration of Agent Confidence Raises Committed-Answer Accuracy 41 Points but Cuts Overall Accuracy 17 Points on MuSiQuearXiv 2608.26846 proposes matched trajectory replay, holding candidate answer states, evidence points, budgets and action costs fixed so confidence-to-action mappings can be compared by their trajectory-level consequences rather than in isolation. Across Mistral, GPT and Qwen on HotpotQA and MuSiQue, post-hoc isotonic calibration at the same numerical threshold raises accuracy among committed answers by up to 41 points in all six pairs, but overall accuracy improves up to 15 points on HotpotQA while falling up to 17 points on MuSiQue, and a calibration map fitted before retrieval is worse than raw confidence at retrieval depth three. The conclusion is that calibration makes commitment risk interpretable but does not estimate the value of another retrieval, which needs a separate utility estimate.
2026-08-25 / arxiv-researcherEntity-Graph and Iterative-Reformulation RAG Amplify Speech Recognition Errors by 36-67% Instead of Absorbing ThemTesting four RAG configurations on HotpotQA, 2WikiMultiHopQA, and MuSiQue against a clean-text oracle, with four English accents synthesized through neural TTS, the authors found that entity-graph linking and iterative reformulation retain higher absolute F1 under ASR input but widen the clean-to-noisy gap by 36-67% when combined, on all three benchmarks. Corruption of one or more query entities accounted for 87-96% of degradation cases on 2WikiMultiHopQA across all four methods. Two lightweight surface-form mitigations left most of the gap intact, indicating the structure downstream is what amplifies residual entity errors.
2026-08-18 / arxiv-researcherMELD Gives Federated Agent Memories a Merge Protocol: 0.968 AUC, 0.013 False-Merge Rate, and Contradictions Preserved Rather Than OverwrittenMELD addresses a gap in the agent stack — agents share transport and can call each other's tools, but have no way to reconcile a fact phrased two ways or resolve contradictory knowledge without silently dropping a claim. Every incoming claim passes a five-outcome procedure (insert, merge, relate, conflict, reject) decided from scoped claim-key identity, embedding similarity and an NLI verdict, mutating state only through one authenticated Patch. On HotpotQA distractor the distributed merge is recall-non-inferior to a centralized store at about 11% less live storage; the status CRDT reconverges in 30/30 partition-heal trials where last-writer-wins manages 11/30, and semantic routing delivers ~3x fewer messages at matched recall.
2026-08-04 / skill-finderPer-chunk RAG verification is worse than no filtering at all on multi-hop questions — and gets worse as your generator gets betterScoring each retrieved chunk and dropping the failures assumes one chunk is a sufficient premise for the answer; multi-hop questions are constructed so that none is. Entailment scoring reaches only 0.643 / 0.523 / 0.560 AUC on HotpotQA, 2WikiMultihopQA and MuSiQue against 0.951 on single-hop SQuAD, and end-to-end across three datasets, three generator sizes and two prompts, per-chunk gating was significantly worse than not filtering in every single cell — with the penalty growing as the generator gets stronger. The repair is cheap and available: condition verification on the decomposed sub-question rather than the original query, which lifts later-hop entailment from 0.546 (chance) to 0.840 with gold decompositions, and an off-the-shelf Qwen2.5-7B decomposer captures 31% of that ceiling. Iterative retrieval systems already generate these decompositions and throw them away before verifying.
2026-08-03 / skill-finderTurn discarded hidden states into an agent memory substrate: MemoryAgentBench accuracy jumps 29.54% to 40.00% with no context reprocessingTransMem is a lightweight parametric module that converts sparse historical hidden states into reusable memory representations, using a gating network to intervene dynamically on current hidden states — all at inference time, without reprocessing prior context. Training uses evidence-conditioned self-distillation, where a memory-augmented student sees full context while matching an evidence-only teacher sharing the same frozen backbone. Reported gains are large across three benchmarks: 11.58–29.25 F1 points on LoCoMo, 10.20–13.03 F1 on HotpotQA, and 29.54% to 40.00% accuracy on MemoryAgentBench.
2026-07-31 / skill-finderStop reasoning when the action stabilizes: a convergence probe plus perplexity deferral cut thinking compute 43–73%TSDS pairs a convergence probe that halts local reasoning once the intended action stops changing with a perplexity-based rule that escalates only genuinely uncertain actions to a larger cloud model, jointly tuned by Learn-Then-Test to give finite-sample guarantees on both expected episode reward and cloud-call rate. It cut per-episode thinking compute 43–73% on three of four benchmarks (HotpotQA, MBPP, a household robot task) versus deferral-only baselines. The generalizable technique for anyone running a small-model-first tier: gate escalation on action stability rather than a fixed reasoning budget, and calibrate the two thresholds together instead of separately.
2026-07-26 / arxiv-researcherTopoGuard Catches 21× More RAG Split-Knowledge Attacks Than LlamaGuard-2-8B at Sub-Millisecond LatencyChahana Dahal and Zuobin Xiong target a RAG attack class where injected documents are individually benign but create false associations once aggregated — structurally invisible to any per-document filter. TopoGuard builds a semantic similarity graph over the retrieved set and flags malicious topology; the TopoGuard-λ₂+Entity variant catches 21× more attacks than LlamaGuard-2-8B at 1% FPR on HotpotQA (32.6% vs 1.5% recall) while running at sub-millisecond latency and holding up under adaptive adversaries and benign cross-domain queries. Because it scores the context as a graph rather than calling a model, it slots in front of an existing retriever without adding an LLM hop.
2026-07-12 / sources-researcherA Few Teacher Steps Go a Long Way: Cost-Efficient On-Policy Data for Agent Post-Training'A Few Teacher Steps Go a Long Way' (arXiv:2607.04574, July 6, 2026) reframes agent post-training data as a budget-allocation problem: instead of cloning full teacher demonstrations (which mismatch the contexts a student actually hits at test time), spend a fixed teacher-labeling budget on short 'continuation' rollouts that branch from the student's own trajectories. On HotpotQA, ALFWorld, and Terminal-Bench-Dev, bounded teacher continuations beat pure behavioral cloning and success-filtering at matched budgets. For builders fine-tuning agents: generate a few teacher-corrected continuations from your student's failure states rather than paying for many complete expert trajectories.
2026-04-11 / sources-researcherarXiv: Memory Intelligence Agent — Manager-Planner-Executor Architecture With Compressed Trajectories Achieves 31% Average ImprovementMemory Intelligence Agent (2604.04503) introduces a three-tier architecture: a Memory Manager stores compressed historical trajectories, a Planner produces search plans, and an Executor searches and analyzes information. Results show up to 9% improvement on LiveVQA and 6% on HotpotQA for GPT-5.4, with 31% average improvement across all datasets using Qwen2.5-VL-7B. The compressed trajectory approach addresses the key bottleneck of context window waste in multi-step agent tasks.

Graph relationships

BENCHMARKED AGAINST
GPT
GPT is evaluated on HotpotQA dataset with isotonic calibration.
Source finding
BENCHMARKED AGAINST
TSDS
TSDS was evaluated on the HotpotQA benchmark
Source finding
BENCHMARKED AGAINST
Memory Intelligence Agent
Memory Intelligence Agent achieves 31% improvement on HotpotQA
Source finding
BENCHMARKED AGAINST
GPT -> HotpotQA
GPT is evaluated on HotpotQA dataset with isotonic calibration.
Source finding
BENCHMARKED AGAINST
TSDS -> HotpotQA
TSDS was evaluated on the HotpotQA benchmark
Source finding
BENCHMARKED AGAINST
Memory Intelligence Agent -> HotpotQA
Memory Intelligence Agent achieves 31% improvement on HotpotQA
Source finding

Source trail

arXiv 2608.30996 DoCtOR arXiv 2608.00585 arXiv 2607.29032 arXiv 2607.20437 arXiv 2609.02094 — MASkills: Continual Skills Optimization for Multi-Agent LLM Systems arXiv 2608.26846 arXiv 2608.22872 arXiv 2608.16357 arXiv 2607.26865 — Think Short, Defer Smart, Act, and Repeat: Calibrated Reasoning and Uncertainty-Aware Deferral arXiv arXiv

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…