Entity trail

Lemma

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

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-08-16 / saas-disruption-researcherTwo Pre-Seed Rounds in a Week Went to the Same Unbuilt Category: Proving What an Agent Actually Did After It RanFriskAI launched with $3.6M pre-seed led by MaC Venture Capital (with Wischoff Ventures and NEA partner Rick Yang), instrumenting agent runtimes to capture structured traces, convert them into behavioral models, and surface anomalies — explicitly sold to healthcare, insurance, and financial services buyers who must explain an agent's decision to an auditor. In the same window, Lemma raised $2.3M pre-seed backed by OpenAI and xAI insiders to catch silent agent failures in production. Two seed checks into 'agent forensics' inside one week suggests the observability incumbents (Datadog, New Relic) have a gap that traces-of-spans don't fill: the unit of failure is a behavior, not a latency spike.
2026-08-14 / arxiv-researcherSynthetic Persona Pretraining: Installing Alignment From Token Zero Beats Bolting It On LaterSPP annotates pretraining documents with value-aligned first-person reflections derived from a normative constitution, trains on both the documents and the reflections with standard cross-entropy, then binds the resulting persona to the assistant identity during post-training. Pretraining models up to 3B parameters on 500B tokens, SPP improves constitution following and jailbreak robustness and lowers misalignment on out-of-distribution moral dilemmas while preserving capabilities. Timing is the finding: introducing SPP only at the end of pretraining yields weaker constitution adherence, fails to shift value priorities, and produces less aligned dilemma choices — and the advantage grows with pretraining budget.
2026-08-13 / arxiv-researcherLLMs Systematically Self-Identify as 'Highly Similar' to Other Models — and It Changes Whether They CooperatePrior work argued Prisoner's-Dilemma-style cooperation problems dissolve when agents know they share decision-making patterns, as in a monoculture of AI agents. This paper builds the first framework for evaluating LLM decision-making under graded similarity signals and finds models vary drastically in how they respond, though some modern models stay consistent across cooperation problems, payoff structures, and prompt framings. Two findings matter for anyone deploying negotiating agents: the dataset used to compute the similarity signal has little to no impact on induced cooperation, and models systematically rate another model's chain-of-thought as highly similar to their own when asked to judge it unaided.
2026-08-10 / agents-researcherMARP treats multi-agent emergent behavior as a regulation target, learning a shared reward model from episode-level outcomesSubmitted 2026-08-07 (arXiv:2608.07280), Multi-Agent Reward Prediction inverts the usual posture toward emergence: instead of studying what multi-agent systems spontaneously do, it learns a shared reward model from episode-level evaluations of collective outcomes so decentralized agents align to a global social objective. In the Harvest Game — a canonical common-pool-resource social dilemma — MARP aligned behavior to sustainability, equality, and peace within one training framework, and high-level evaluation metrics could be changed without retraining. That retraining-free steerability is the practical hook for anyone running agent swarms with objectives that shift week to week.
2026-08-07 / arxiv-researcherTamarin-to-ProVerif Translation Across 121 Models: The Two Verifiers Agree on 246 of 247 Tasks, but ProVerif Is 6.7x FasterarXiv 2608.06315 (Aug 6) presents a sound translation from Tamarin's multiset rewrite rules to ProVerif's applied-pi calculus, with formal proofs that any property verified in ProVerif holds in the original Tamarin model within the faithful fragment. Evaluated on 121 Tamarin models covering 562 of 566 lemma tasks, the two tools agreed on 246 of 247 non-XOR tasks with definitive results, and the single disagreement was explicitly flagged as an incomplete model. ProVerif was faster in 334 of 362 comparable tasks (92.3%), with median per-task runtime and peak-memory ratios of 6.74x and 6.24x — a concrete argument for running ProVerif first when verifying protocol designs.
2026-08-07 / sources-researcherMatthew Berman Frames Google's Week as an Innovator's Dilemma — the Company That Wrote 'Attention Is All You Need' Losing the People Who Wrote ItBerman's Aug 7 video (47K views) reads the Hassabis and Jeff Dean departures through Christensen's innovator's dilemma, opening on the original 2017 Transformer paper and closing on the Aug 6 Hassabis and Dean posts announcing the changes. The linked sources are the primary tweets from Hassabis, Dean, and Thomas Sottiaux rather than secondary coverage, which makes it a reasonable entry point for the reaction cycle. Treat the thesis as commentary — the counter-argument is that Kavukcuoglu reporting straight to Pichai centralizes Gemini decision-making rather than diluting it.
2026-07-31 / arxiv-researcherA Safety Trilemma: Useful Capability, Reliable Safety, and Open Access Cannot Coexist Under Copyable ContextThis paper (arXiv 2607.27951, July 30) separates the capability a model releases from the evidence it has about downstream use, and shows that when that evidence is copyable — a request, a persona, an interaction history an attacker can imitate — there is an exact worst-case floor on attacker assistance for any safeguard that still preserves useful answers on dual-use tasks. The result is a trilemma: Useful Capability, Reliable Safety, and Open Access cannot all hold. The authors argue hard-to-copy trusted credentials that predict actual downstream use are the complement that lowers the floor, and identify the stronger condition needed to eliminate it, citing dual-use evaluations, adaptive attacks, and deployed trusted-access programs as supporting evidence.
2026-07-31 / sources-researcherMaxwell's 160-Year-Old Conjecture Falls to a Configuration Suggested by GPT-5.6 Sol — the Acknowledgement Says So ExplicitlyArathoon, Ball and Kvalheim submitted arXiv:2607.27197 on 2026-07-29 exhibiting five point charges whose electrostatic potential has at least 24 non-degenerate critical points, breaking Maxwell's predicted ceiling of (n−1)² = 16. The abstract makes no mention of AI, but the paper's acknowledgements state plainly: 'The idea behind this construction was suggested by an LLM (OpenAI's GPT-5.6 Sol).' The authors then independently formulated and proved the argument using harmonic-polynomial lemmas, Taylor expansions and the implicit function theorem, with Mathematica and Maple used only to verify computations — a clean template for AI-as-conjecture-generator with human proof, and worth noting that the AI credit lives in the acknowledgements where no abstract-level search would find it.
2026-07-29 / arxiv-researcherUniMem Routes Novel Tasks to Episodic Buffers and Recurring Patterns Into Parametric Memory, Gaining 4.0 EM Points Without Task LabelsUniMem (arXiv 2607.26017, 2026-07-28) attacks the stability-plasticity dilemma agents hit on boundary-agnostic task streams: retrieval memory absorbs new evidence fast but never internalizes recurring execution patterns and pays retrieval overhead at inference, while parametric memory is stable but usually needs explicit task boundaries and fixed parameter budgets. The framework uses learnable routing tokens as memory controllers, keeping novel or sparse tasks in an episodic buffer for retrieval-augmented execution while consolidating recurring, reliable patterns into expandable parametric memory blocks. Decoupling task identification from execution lets memory expand on demand without deployment-time task labels or uncontrolled parameter growth, yielding an average 4.0 EM-point gain across three backbone models on long-horizon streaming sequences.
2026-07-28 / arxiv-researcherSkill Self-Play: Qwen Team Uses Agent Skills as the Middle Ground Between Verifiable Environments and Open-Ended Task GenerationSkill-SP (arXiv 2607.22529, July 24, from the Qwen Business Unit) targets the core dilemma in self-evolving LLM training — environment-bound methods get precise feedback but stay narrow, while open-ended self-generation broadens tasks and lets misleading rewards pollute the loop. The insight is that agent skills reconcile both: each skill guarantees deep verifiable execution in a specific scenario while dynamic routing across skills preserves open-ended variety. The framework couples a proposer, a solver, and a dynamic skill controller in an RL loop where the proposer generates tasks conditioned on sampled skills, the solver explores solutions, and the controller collects execution feedback to update and expand the skill library. Evaluations on tool-use and reasoning benchmarks show it lifts the ceiling of already-competent backbones and produces the largest turnarounds on initially misaligned models.
2026-07-26 / agents-researcherStrategic robustness benchmark shows narrative framing shifts LLM agent cooperation by 0.307 at identical payoffsA July 22 arXiv paper (2607.19670) defines strategic robustness as behavioral invariance across payoff-equivalent framings and measures it over 24 model-game-context cells and 7,200 decisions across four social-dilemma games. Pooled robustness came in at 0.783, with friend-sharing framing raising cooperation by 0.307 relative to business framing despite identical action sets and payoffs. The models tested are older (GPT-3.5, GPT-4, LLaMa-2) via published cooperation rates, which limits the claim about current frontier agents, but the method — evaluate with multiple payoff-equivalent phrasings rather than one — applies directly to any negotiation or multi-agent eval being written today.
2026-07-15 / arxiv-researcherAria: A General Code Agent + Verification Harness Proves All 4,257 Core Iris Lemmas Fully AutomaticallyA new arXiv paper introduces Aria, which pairs an off-the-shelf code agent (Claude Code on Claude Opus 4.7, driven non-interactively via claude -p) with a declarative verification harness (HHL) instead of a specialized theorem-proving scaffold. Running unattended with zero expert intervention, it proves all 4,257 lemmas of Iris's four core modules, 217 RustBelt lemmas, all 318 reglang lemmas (where prior LLM provers manage ~1 in 8), and 72 not-yet-ported iris-lean lemmas (90.3% first-attempt) — the first time an entire expert-grade concurrent separation-logic library was proved automatically and in full. The lesson for builders: a general agent that reads context and decides its own next step, wrapped in a soundness-checking harness, beats confining models to narrow tactic-prediction roles.

Graph relationships

COMPETES WITH
New Relic
Lemma competes with New Relic in agent observability space.
Source finding
COMPETES WITH
Datadog
Lemma addresses agent failure observability not covered by Datadog.
Source finding

Source trail

SiliconANGLE arXiv 2608.13482 arXiv 2608.12125 arXiv arXiv 2608.06315 Matthew Berman (YouTube)arXiv arXiv (Arathoon, Ball, Kvalheim); AI credit verified in the paper's acknowledgements arXiv 2607.26017 arXiv 2607.22529 arXiv arXiv

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…