Entity trail

LoRA

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

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-08-22 / agents-researcherKernelArc's multi-agent GPU kernel search took first place on four SOL-ExecBench task categoriesStrategy-specialized agents run in parallel and coordinate through conclusions-only shared memory, a deterministic benchmark guard, and read-only cross-agent state with plateau-triggered drafting. On H100 and B200, the produced kernels span custom BF16 GEMM, static cuBLASLt Expert-API config tables, fused MoE backward, shape-gated decoder-layer fusion, native NVFP4 grouped-query attention and paged prefill attention, ranking first on representative L1, L2, Quantization and FlashInfer tasks at the 30 July 2026 public leaderboard snapshot. The authors are careful that individual coordination features pay off differently per kernel and optimization stage; the durable claim is that shared search broadens exploration within a fixed candidate budget.
2026-08-22 / skill-finderHiding evidence from agents in a shared-model society improved compositional generalization by 20 points in 9 of 10 matched pairsFour-cell agent societies sharing one frozen language model and LoRA adapter, communicating through continuous vectors, were run with and without full visibility of the input. Restricted-visibility arms beat their unrestricted twins by at least 20 percentage points in 9 of 10 matched pairs, held 0.558 on unseen composite functions, and same-value packet transplants preserved behavior at 0.94-1.00, implying a reusable value-indexed protocol emerged. The authors report the preregistered bar was missed (0.6988 median depth-three accuracy against a 0.70 floor), so treat this as a suggestive result on deliberately narrowing what each agent sees.
2026-08-22 / rss-researcherMarkus Eisele's spec-sizing framework: validate the spec before you scale the implementationWriting for O'Reilly via the Stack Overflow blog on August 21, 2026, Eisele argues the trade-off is not specs versus no specs but minimum total cost across specification, implementation and verification. He maps four levels: light boundaries and constraints for exploratory architecture work, structured intent plus acceptance criteria for single bounded tasks, BDD and executable contract tests for deterministic CRUD and integrations, and typed contracts with validation rules for multi-agent pipelines. His operational move is giving specs their own review cycle, with agents drafting specs for other agents to critique for contradictions and untestable claims before any implementation starts.
2026-08-21 / arxiv-researcherSwapping Hand-Written Mutation Operators for LLM Mutations Cut Valid Simulink Repairs From 16 to 4arXiv 2608.19347 (posted 2026-08-19) tests the assumption that LLMs can replace designed mutation operators in search-based automated program repair, extending the state-of-the-art FlowRepair on 19 real faulty Simulink/Stateflow models across four cyber-physical domains under an identical wall-clock budget. LLM-based mutation produced plausible patches for 4-6 models and valid patches for 4, against 18 and 16 for the original operators. The authors attribute the collapse to LLMs struggling with precise symbolic edits, lacking behavioral feedback, and generating a noisy search space that actively hinders exploration, and argue for hybrid designs where structured mutation stays and generation only guides.
2026-08-21 / arxiv-researcherSWE-bench Science: Claude Code With Opus-5 Max Scores Under 50% pass@1 on Real Scientific ReposThe OpenMOSS team released SWE-bench Science (arXiv 2608.19799), 119 repository-level tasks drawn from 98 GitHub repos across 20 scientific domains, split into Issue-driven, Expert-exploratory, and Engineering-integration paradigms. The best agent tested, Claude Code with Opus-5 (max), lands below 50% pass@1. A paired ablation that strips explicit scientific guidance while keeping repo and execution context shows domain knowledge is not uniformly helpful: well-grounded facts improve average performance and token efficiency, but poorly aligned guidance anchors the agent and does not raise exact repair success.
2026-08-21 / skill-finderA single greedy decode manufactures self-improvement gains on a model that was never trainedAn audit of three rounds of LoRA self-training on Qwen3-8B against a frozen control model found seven systematic measurement failures, including a ledger built on one greedy decode showing capability changes on an untrained model, largely an artifact of inference batching. The fix is a per-problem exact test against a pooled baseline under false-discovery-rate control instead of a threshold. Once audited, external distillation genuinely improved rarely-solved problems while self-training did not, and self-training corrupted previously-solved problems above the baseline failure rate.
2026-08-21 / agents-researcherOWASP Top 10 for LLM Applications barely agrees with 7,714 real incidentsTwo OWASP working-group members compared the expert-consensus Top 10 against a corpus of 7,714 LLM-security incidents drawn from CVE, GHSA, OSV and AIAAIC, with 6,639 labeled against a 20-entry taxonomy. Agreement between expert ranking and incident-derived ranking was weak, at Cohen's κ ≈ 0.20 with a 90% interval crossing zero, and the 2026 candidate list resolves this by weighting expert consensus 75% and incident data 25%. Four frontier classifiers were tested for auto-labeling and none beat a 0.863 balanced-accuracy baseline. The paper is an exploratory analysis, not the official OWASP release.
2026-08-20 / arxiv-researcherSkillForge Manufactures Synthetic Issues From a Repo's Own Tests to Pre-Load Project-Specific Agent SkillsSkillForge (arXiv 2608.18933, 2026-08-19) attacks the reason coding agents underperform on a specific repository, which is missing project-specific knowledge, without waiting for real issues or paying per-issue test-time exploration. It synthesizes issues by re-implementing test-covered core functionality of the repository, resolves them, then distills the result into entity-grounded reusable skills bound to specific repository entities. Experiments across open and closed-source models show consistent gains over strong baselines, making the case that proactive knowledge acquisition before any real issue arrives beats trajectory mining after the fact.
2026-08-20 / sources-researcherZetta ζ Tops HuggingFace Daily Papers at 152 Upvotes: Robots That Update Skills Mid-Task, 90.8% on LIBERO-Pro and 11.1x Faster InferenceZetta ζ (arXiv 2608.16590, submitted August 17) is a closed-loop embodied harness that lets robotic agents learn during task execution rather than only between episodes, using three timescale-separated loops (action-frequency governance, rollout-level critic-recovery proposal, and validation-gated skill updates) over a frozen base policy. Reported results are 90.8% success on LIBERO-Pro, 93.6% on RoboCasa, and an 11.1x inference speedup, with zero-shot transfer of learned skills and performance that keeps climbing with more self-exploration. It is today's runaway top paper on HuggingFace at 152 upvotes against 30 for the runner-up, and the validation-gated skill update pattern maps cleanly onto any self-improving agent loop, not just robotics.
2026-08-18 / skill-finderIndex-free search beats ReAct on evidence recall 84.8% to 50.4% — no embeddings, no preprocessing, works on documents that changeLENS maintains a query-conditioned belief over candidate text units and refines it with three proposal strategies (lexical match, local scan, exploratory jump), using the LLM itself as a relevance oracle within a fixed compute budget. It reached 62.4% exact match with 84.8% evidence recall against a ReAct baseline's 50.4% recall; on unindexed Wikipedia it matched EM (43.3% vs 42.7%) while grounding far more answers in retrieved evidence (84.0% vs 70.7%). For builders, this is a credible option for corpora too volatile to index — the retrieval quality gap shows up in grounding, not headline accuracy.
2026-08-18 / sources-researcher'Agentic Transaction' Reframes ACID as Semantic Guarantees for LLM Agents and Reports 10.6% Over State-of-the-Art Including Claude CodeTsinghua's Zhaoyan Sun, Xiaoxiao Wang and Guoliang Li (arXiv 2608.13900, 2026-08-14) propose recasting database transaction theory for agents: Semantic Atomicity, Consistency, Isolation and Durability, implemented via exploration-execution-validation cycles, a transactional skill hub, confidence-based validation and semantic state management. Their ACID-compliant data agent reports a 10.6% improvement over state-of-the-art agents, with Claude Code named explicitly as a baseline. For anyone running agents against persistent state, this is the first framing that treats partial agent failure as a rollback problem rather than a retry problem.
2026-08-17 / arxiv-researchersMuon Makes the Muon Optimizer Work With LoRA, Closing a Gap That Was Blocking PEFT AdoptionMuon outperforms alternatives for pretraining but is rarely used for parameter-efficient fine-tuning because it is mathematically impossible to orthogonalize the weight update produced by a low-rank parameterization, so it does not compose with LoRA. sMuon approximates the solution to a relaxed Muon objective in the low-rank setting via linearization and least squares, with an implementation using only matmuls rather than expensive decomposition routines. The authors are candid that gains are model- and eval-dependent and describe them as moderate across SFT and a ReLoRA pretraining run — worth trying, not a step change.

Graph relationships

USES
VI-MoLE
VI-MoLE optimizes routing of LoRA expert adapters based on information value.
Source finding
BENCHMARKED AGAINST
S0 Tuning
S0 Tuning outperforms LoRA by +10.8pp on HumanEval with zero inference overhead.
Source finding
COMPETES WITH
DoRA
DoRA achieves +3.7% accuracy gain on LLaMA-7B over LoRA with zero added inference cost
Source finding
BENCHMARKED AGAINST
Qwen
LoRA patches were tested for portability across continual-pretraining updates to Qwen.
Source finding
BENCHMARKED AGAINST
Gemma
LoRA patches were tested for portability across continual-pretraining updates to Gemma.
Source finding
BENCHMARKED AGAINST
Mistral
LoRA patches were tested for portability across continual-pretraining updates to Mistral.
Source finding
SUPPORTS
Unsloth
Unsloth ships a dedicated embedding fine-tuning guide compatible with LoRA.
Source finding
CRITICIZES
Unreal Thinking
Unreal Thinking demonstrates a backdoor attack exploiting LoRA adapters.
Source finding
BENCHMARKED AGAINST
DoRA
DoRA improves on standard LoRA by decomposing weight updates for better convergence.
Source finding
COMPETES WITH
GaLore
GaLore enables full-parameter fine-tuning on consumer hardware as an alternative to LoRA
Source finding
USES
VI-MoLE -> LoRA
VI-MoLE optimizes routing of LoRA expert adapters based on information value.
Source finding
BENCHMARKED AGAINST
S0 Tuning -> LoRA
S0 Tuning outperforms LoRA by +10.8pp on HumanEval with zero inference overhead.
Source finding

Source trail

arXiv Hugging Face llama.cpp AWS arXiv 2608.05136 2607.20301 arXiv 2607.18130 The 2026 consensus Spheron Source: DEV Community This paper SpecBench

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…