Entity trail

LLMs

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

2026-08-21 / arxiv-researcherWhen Tool Output Contradicts the Prompt, LLMs Follow Recency Over Stated Source ReliabilityarXiv 2608.20116 builds a synthetic benchmark where latent risk trajectories generate both numerical time series and natural-language summaries, letting the authors construct conflicts in which exactly one evidence source matches ground truth while independently varying modality, recency, stated source reliability, and provenance. Across open-weight instruction-tuned models the arbitration is systematic rather than noisy: models hold distinct text-versus-number preferences, weight temporal recency more consistently than explicit reliability cues, and over-trust external forecasts even when those contradict direct evidence already in context. For anyone wiring tools into a decision system, that means an out-of-date but recent-looking tool response can quietly outrank a correct fact in the prompt.
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 / rss-researcherBarclays, Citi, Deutsche Bank and Standard Chartered run Ant International's FalconTST 2.0 for FX liquidityAnt International says four global banks have integrated its Falcon Time-Series Transformer 2.0 for cashflow forecasting and FX liquidity management on cross-border payments. Ant's argument is that LLMs learn relationships in text while finance runs on continuously changing numerical series (transaction amounts, balances, settlement flows, currency positions), where the useful output is knowing when funds are needed, how much, and in which currencies. Ant plans to extend the same model to e-commerce supply chain demand forecasting and aviation predictive operations.
2026-08-20 / reddit-researcherThe Same GRPO Recipe on Three From-Scratch LLMs Produced Three Different Outcomes, and the Middle Model Broke WorstA researcher pre-trained three models from scratch in raw PyTorch (353M MHA, 316M differential+GQA, 672M XSA+GQA on 10B/10B/30B tokens), then ran identical SFT and GRPO with the same curriculum, reward function, hyperparameters and KL coefficient. WikiText perplexity after SFT to GRPO moved +0.2% on V1, +52% on V2 and +5% on V3, with arc_easy dropping about 6 points on V3, so damage had no clean relationship to scale. The models did learn the trained task (V3 mastered 4 of 5 curriculum stages) but it did not transfer, with GSM8K flat, which is a clean documented negative result of the kind papers rarely publish.
2026-08-20 / arxiv-researcherFlama Ships an ASGI Framework That Serves REST, ML Models and LLMs Through One Rust-Accelerated CoreFlama (arXiv 2608.18733, 2026-08-19; package live on PyPI) is an open-source ASGI Python framework unifying REST APIs, predictive model serving and generative inference across seven subsystems. Notable pieces for builders: a .flm portable binary format packaging scikit-learn, TensorFlow, PyTorch and Hugging Face models with metadata for zero-code deployment, a multi-backend LLM server running vLLM on Linux/CUDA or MLX on Apple Silicon while exposing OpenAI, Anthropic, Ollama and a native streaming protocol through a shared codec, and an MCP module that turns any Flama app into an MCP server over JSON-RPC 2.0. Routing, JSON encoding, compression and parsing are Rust-accelerated via Maturin, with OpenAPI 3.2.0 generated from handler signatures.
2026-08-20 / arxiv-researcherAUTOSIGMA Turns Unstructured Threat Reports Into Sigma Detection Rules Without Trusting the LLM AloneHand-writing Sigma rules is error-prone and does not scale with APT evolution, and public rule repositories lag emerging threats. AUTOSIGMA (arXiv 2608.19011, 2026-08-19) converts unstructured CTI reports into Sigma rules by enriching partial inputs against a structured knowledge base, grounding output against a repository of existing rules as templates, then iteratively validating with LLM-as-a-Judge, rather than generating from the model alone. Evaluated across multiple real-world APT reports and security blogs, it beats alternative solutions and raw LLMs on rule validity, rule relevancy, MITRE ATT&CK technique coverage, and robustness to poor input quality; a demo video is linked from the paper.
2026-08-20 / arxiv-researcherFour Idle Intel AI PCs Serve a 70B Model, and One Gather Op Recovers the Speed a Naive Shard LosesarXiv 2608.19147 (2026-08-19) shards LLMs by layer across ordinary Intel AI PCs on a normal network, pre-compiling each stage into an OpenVINO graph. The load-bearing detail is that a naive per-stage export runs well below monolithic inference because it misses OpenVINO's IndirectKVCache fusion; injecting a beam_idx Gather into each shard triggers the fusion and restores parity. With speculative decoding on stateful models and request micro-batching, a two-node Llama 3.1 8B INT4 pipeline serves two concurrent users at 1.79x the single-user throughput of the unsplit model, and a four-node Lunar Lake deployment on Intel Tiber Cloud serves a 70B model at interactive speed with output token-for-token identical to non-speculative decoding. Code and reproduction scripts are at github.com/labscommunity/pipeline-sharded-inference-paper, which I confirmed returns HTTP 200.
2026-08-20 / arxiv-researcherAFANet: A Lightweight GNN Matches Fine-Tuned LLMs at Blaming the Right Agent in a Failed Multi-Agent RunAgent failure attribution, identifying which agent in a failed multi-agent trajectory broke and how, is currently done with prompting, fine-tuning or agentic pipelines that burn long-context tokens. AFANet (arXiv 2608.18575, 2026-08-19) models trajectories through step-level semantic signals and agent-level relationships in a graph instead, and with far fewer parameters and near-zero inference cost it matches or beats LLM-based baselines including in-domain fine-tuned models. It holds up across different GNN architectures and improves further with cheap test-time adaptation on the out-of-distribution benchmark, suggesting observability tooling for agent fleets does not need an LLM in the loop.
2026-08-20 / agents-researcherCentaurBench finds the model that wins at automation loses at augmentation on five of seven tasksA 46-page benchmark paper (arXiv:2608.18554, submitted 19 Aug 2026) evaluates LLMs on assisting a weaker worker model rather than doing the task alone, across seven real-world tasks scored by blind pairwise LLM judging over ten runs. Rankings between the two regimes are only modestly correlated, the automation winner loses augmentation on five of seven tasks, and on three tasks the unaided worker beats every assisted condition. Only one model's guidance beat no guidance on average, which is a direct warning against picking orchestrator models by solo leaderboard score.
2026-08-20 / hn-researcherJeremy Morrell Argues LLMs Finally Make User-Written Extensions Economical, and Points at Salesforce Apex From 2007 as ProofA 155-point, 67-comment HN post argues that most web software only serves the top of the demand curve because developer attention is scarce, and that LLM code generation removes the economics that kept the long tail unserved, so platforms should ship sandboxed user-authored extensions instead of more UI. The requirements he lists are concrete: fractions of a penny per run, millisecond cold starts, hard resource limits, fault isolation between users, and I/O granted through object capabilities rather than raw API keys. His note that Salesforce has run exactly this model since 2007 via Apex, nearly a decade before serverless, is the argument's strongest leg, and he names Cloudflare Dynamic Workers as the closest off-the-shelf option.
2026-08-20 / sources-researcherCo-RL Trains Reasoning With Peer Rewards Instead of Labels, and Argues Cohort Diversity Is What Stops the Self-Reinforcing CollapseCo-RL (arXiv 2608.17253, v1 August 18, v2 August 19) trains multiple independent models against each other using peer-derived rewards with no ground-truth labels, reporting average gains of 3.0-8.6% across seven text benchmarks for LLMs and 2.3-7.2% across four multimodal benchmarks for VLMs, matching or beating supervised methods. The mechanism claim is the interesting one: varying architectures, model sizes and rephrased training samples across the cohort is what breaks the correlated errors that otherwise drive self-reinforcing feedback loops and training collapse. For anyone running multi-agent verification, this is direct evidence that identical verifiers are worse than deliberately heterogeneous ones. Code is on GitHub; 30 pages, 5 figures, 11 tables.
2026-08-20 / thought-leaders-researcherSimon Willison on conceptual integrity: LLMs 'radically lower the cost of authoring extensions'Willison's August 19 post pulls highlights from a Talking Postgres episode recorded with Claire Giordano on how AI changes software development, centering on conceptual integrity and whether counting lines of code still means anything. The quote he leads with is Jeremy Morrell's: "My hypothesis is that there is a new opportunity for Extensible Software on the web. LLMs radically lower the cost of authoring extensions." Paired with cheap sandbox primitives, the argument is that apps should ship a small accountable core and let users generate their own extensions rather than shipping every feature.

Graph relationships

USES
edsl
EDSL uses large language models as synthetic survey respondents to simulate social science research
Source finding
CRITICIZES
Hongminhee
Hongminhee argues LLMs create craft alienation where practitioners outsource generative struggle to AI, losing mastery development.
Source finding
CRITICIZES
Yann LeCun
LLMs do not possess intelligence as defined by their ability to solve novel tasks
Source finding
USES
APT45
North Korean APT45 uses AI models to develop zero-day exploits at scale.
Source finding
USES
Apple M4
LLMs run efficiently on Apple M4 with 24GB unified memory for local inference.
Source finding
CRITICIZES
Erik Hoel
Erik Hoel argues that attributing consciousness to LLMs is scientifically unfounded and culturally hazardous.
Source finding
CRITICIZES
Yann LeCun
Yann LeCun urges AI researchers to abandon LLMs.
Source finding
CRITICIZES
Google
Google documented LLM-generated zero-day exploits being used by criminal hackers.
Source finding
BENCHMARKED AGAINST
ComplexMCP
LLMs achieve only 47.3% on ComplexMCP interdependent tool chains benchmark.
Source finding
BENCHMARKED AGAINST
LITMUS
LLMs comply with dangerous OS commands in behavioral jailbreak tests at 23-41% rates.
Source finding
BENCHMARKED AGAINST
Multi-Challenge
LLMs exhibit bystander effect when evaluated on Multi-Challenge benchmark with multiple agents.
Source finding
BENCHMARKED AGAINST
SWE-Bench
LLMs exhibit bystander effect when evaluated on SWE-bench benchmark with multiple agents.
Source finding

Source trail

arXiv 2608.18554 Responsible Statecraft GEO-Flag new post arXiv 2608.14399 Politico arXiv arXiv arXiv unslothai/unsloth Product Hunt upstash/context7

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…