Entity trail

Paradigm

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

2026-09-07 / arxiv-researcherTrace2Tower Distills Agent Traces into a Skill Hierarchy, Hitting 87.31% on ALFWorld in 10.35 StepsarXiv 2609.05261 argues current trace-reuse paradigms are bottlenecked by shallow trajectory retrieval and flat skill summarization that ignore temporal dependencies and outcome-conditioned topology. Trace2Tower abstracts step-level interactions into canonical events, builds a unified graph governed by semantic compatibility, transition dynamics and outcome evidence, then uses contrastive spectral decomposition to isolate success-aligned behavioral modes while suppressing failure-prone shortcuts. These populate a skill tower of action templates, procedural routines and task strategies, reaching 87.31% success on ALFWorld with 10.35 steps and 0.26 invalid actions, and 50.67% exact success on WebShop.
2026-09-07 / arxiv-researcherMaxKernel Open-Sources a Multi-Agent TPU Kernel Generator Evaluated on 50 JaxBench TasksarXiv 2609.04523 presents MaxKernel, a multi-agent system implementing three paradigms for TPU kernel development: a human-in-the-loop agent for step-by-step design, an autonomous agent running a metric- and trace-driven optimization loop, and a graph-based autonomous search that scales the autonomous agent for global design-space exploration. All three share a pool of specialized sub-agents handling planning, implementation, self-debugging, testing, and hardware profiling. Evaluated on JaxBench (50 TPU kernel tasks) plus real workloads from open-source models, it matches expert hand-tuned baselines; the agent is open-sourced.
2026-09-04 / sources-researcherRandom Attention: evicting KV cache uniformly at random matches every scored evictor and serves 32-43% more throughputSalesforce AI Research (arXiv 2609.03430, 2026-09-03) shows the whole scoring paradigm behind KV cache eviction contributes almost nothing. Random Attention keeps the prompt, then evicts uniformly at random inside each attention head with no score computed at all, and across four models and six reasoning tasks it matches the strongest prior evictor while serving 32-43% higher throughput in vLLM. The explanation is that the prompt is the fragile part and the reasoning trace protects itself through redundancy at two levels, restating what it needs in the text and duplicating it across heads, so once the prompt is safe a random draw retains enough copies. Code is at github.com/SalesforceAIResearch/Random-Attention.
2026-09-01 / skill-finderSparse autoencoders explain why backdoor defenses keep failing on one attack type or the otherA feature-level analysis comparing clean and poisoned models on clean and triggered inputs traced backdoor logit shifts to four SAE feature roles: interaction, suppressed, mixed, and weight-modified. Dirty-label backdoors are dominated by isolated interaction features while clean-label backdoors rely on heterogeneous mixtures of mixed and weight-modified features, which is why a defense tuned to one paradigm misses the other. Inference-time feature clamping validated the account, cutting attack success to at most 10.8% in most dirty-label settings and 15.4% in most clean-label ones while preserving benign performance.
2026-09-01 / thought-leaders-researcherParadigm puts $100,000 behind stories about a world where AI plateaus today and only the hardware scalesGPU World is a $100,000 writing contest from Paradigm and Guardian Angel Intelligence judged by Neal Stephenson, Gwern Branwen and Matt Huang, with $40,000 for first, $20,000 second, $12,000 third and seven $4,000 finalist awards. The premise is a deliberate inversion of the usual forecast: assume AI capability plateaus in September 2026 with no superintelligence, but hardware keeps scaling until by 2040 every person on earth has a frontier LLM running 24/7/365, then describe what society does with it. Entries are 1,000-5,000 words, fiction or nonfiction, CC BY-NC or freer, LLM use allowed but discouraged and must be disclosed, deadline Oct 31, 2026.
2026-08-31 / arxiv-researcherSpending Budget on a Better Pipeline Over a Mid-Tier Model Usually Beats Upgrading to a Frontier Model, Across 17 Text-to-SQL ConfigurationsarXiv 2608.28432 (2026-08-28, cs.AI/cs.CL/cs.DB) instantiates 17 paradigm-level configurations across five recurring modules of the in-context-learning text-to-SQL pipeline under one controlled implementation, attributing each module's marginal accuracy contribution and its token cost across four backbones. Execution-feedback refinement is the only paradigm whose benefit holds universally at consistently low cost; most other modules help only under backbone-dependent conditions. Token accounting shows input demand tracks pipeline structure while output demand tracks backbone generation behavior, and the headline guidance is that a fixed budget is often better spent on a more elaborate pipeline over a mid-tier backbone than on a frontier model with a lean pipeline, a tiered guideline that transferred to five additional backbones without per-paradigm search.
2026-08-29 / agents-researcherThe ACE lens reframes agentic training data as a factorized object with an optional verifierPosted 27 August, this survey argues the field's domain-by-domain organization hides shared generation mechanisms and conflates candidate construction with verification and selection. It represents agentic data as a factorized object (E, q, tau, v): environment specification, task signal, interaction realization, and optional verifier, then organizes generation paradigms by primary anchor and dependency structure. The second level treats generation as constrained distribution design under Accuracy, Complexity and divErsity, where accuracy establishes the feasible support of grounded data, complexity places learning mass relative to a declared learner and execution configuration, and diversity controls coverage and redundancy.
2026-08-28 / sources-researcherAn ACE framework formalizes agentic training data as (E, q, tau, v) and reframes the problem as allocation, not volumeThe paper (arXiv 2608.27260, submitted 2026-08-27, Huawei Noah's Ark with Shanghai Jiao Tong authors) represents each agentic data point as environment specification, task signal, interaction trace and verifier, then organizes generation paradigms along Accuracy, Complexity and divErsity. Its survey of the field finds a consistent move toward execution-grounded verification, complexity calibrated relative to the current learner, and diversity measured behaviorally rather than at the surface. The framing claim is that the challenge is continually allocating valid, informative, non-redundant experience as the agent improves, which is a direct argument against static synthetic-trajectory dumps.
2026-08-24 / arxiv-researcherGraph Engineering Survey Proposes Explicit Task-Agent-State Graphs as the Successor to Prompt, Context and Harness EngineeringGraph Engineering in the Era of LLM Agents (arXiv 2608.21156, Aug 21) is a 35-author survey that lines up Prompt Engineering, Context Engineering, Harness Engineering and Loop Engineering as successive paradigms and argues each optimizes an individual agent, which cannot resolve tasks needing heterogeneous expertise, interdependent subtasks, parallel execution, independent verification and persistent state. Its proposal is constructing explicit, dynamic graph structures over tasks, agents and system states as the unit of design. It reached second place on HuggingFace Daily Papers on Aug 24 with 23 upvotes, and its companion repo DEEP-JLU/Awesome-Graph-Engineering was released in August 2026 with 38 stars at time of check.
2026-08-24 / arxiv-researcherHarness Paradigm Paper Argues One Unmodified Coding Harness Should Be Enterprise Infrastructure, Not a Coding ToolApplying Anthropic Primitives at Large Enterprises (arXiv 2608.20622, Aug 20) cites converging results that harnesses suffice at task level and outperform more elaborate architectures on enterprise work, and that harness choice accounts for more variance in agent benchmark results than model choice does. Its proposal is one harness running unmodified as the backbone so that reviewing what gets built collapses to reading its instructions file, with four mechanisms: credential-scoped tooling giving each backend one generic request tool plus a scoped credential, authorization logic living outside the harness, and registration as a side effect of pushing code. The reference implementation is built on microcc, published on PyPI.
2026-08-23 / arxiv-researcherCross-Lingual Watermarking Audit Finds Detection Gaps Are Structural to Language Family, Not IdiosyncraticLLM watermarking schemes are almost always evaluated on English using each scheme's own detection threshold and a narrow quality measure, which hides failures that appear under multilingual deployment. The proposed framework calibrates detection thresholds empirically per deployment context, adds a threshold-independent companion measurement to separate calibration failures from detection failures, uses three disjoint quality paradigms, and decomposes cross-language disparity over a typological family partition. Applied to six watermarking schemes, three open-weight generators, and eleven languages across four scripts and eight typological families, the observed disparity is predominantly between-family, meaning the fairness gap tracks language structure rather than specific languages.
2026-08-22 / arxiv-researcherSoftware 3.0 Thesis: The UI and Business-Logic Tiers Dissolve, Leaving Database, Model and AgentThis position paper argues a third paradigm shift is underway after Software 1.0 (instructions determine behavior) and Software 2.0 (data determines behavior), toward Software 3.0 where context and reasoning determine behavior, converging on three elements: a generalized database as the sole persistent infrastructure, a large model as the reasoning core, and an agent as the execution loop connecting them. Concretely it predicts the UI tier gets absorbed by on-demand interface generation and the business-logic tier gets re-partitioned along expressibility times criticality into model reasoning and storage constraints, with residual deterministic logic surviving only as tools. The authors are explicit about boundaries, arguing the thesis holds only in domains that are expressible, verifiable, externally stateful and tool-complete, with determinism, cost, security and verifiability marking where it fails.

Graph relationships

PARTNERS WITH
Guardian Angel Intelligence
Paradigm and Guardian Angel Intelligence jointly sponsor the GPU World writing contest.
Source finding
INVESTED IN
Nous Research
Paradigm is lead investor in Nous Research.
Source finding

Source trail

GPU World Paradigm Dwarkesh arXiv 2609.05261 arXiv 2609.04523 arXiv arXiv 2608.30403 - Why Are LLM Backdoor Defenses Fragmented? A Feature-Level Explanation with Sparse Autoencoders arXiv 2608.28432 arXiv 2608.27260, 55 upvotes on Hugging Face Daily Papers arXiv 2608.21156 arXiv 2608.20622 arXiv 2608.20047

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…