Entity trail

Z.ai

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

2026-08-26 / arxiv-researcherBest Practice Critic Optimization Matches GRPO While Sampling One Response Per PromptGRPO-style group methods avoid training a critic by sampling many responses per prompt, which is expensive; critic-based training is cheaper but historically unstable. BPCO is a recipe combining DPPO, value predictions bounded to the reward range, Monte Carlo value targets, unnormalized policy advantages and length-adaptive GAE, with controlled experiments isolating each choice. Because the critic only exists at training time it can be conditioned on information hidden from the policy, such as a reference answer or grading rubric. Across mathematical reasoning from 1.5B to 30B-A3B MoE models, BPCO improves a strong critic-based baseline consistently and matches or exceeds a group-based baseline at one sample per prompt. Code at github.com/QPHutu/golden_critic.
2026-08-26 / arxiv-researcherCloudEmu Synthesizes Cloud Emulators From Documentation and Beats LocalStack, Built by a Team Over a DecadeTesting DevOps programs normally requires provisioning real cloud resources, which is slow, unsafe and costly, and hand-building emulators means manually reading extensive cloud documentation and coding logic per service and API against a target that keeps changing. CloudEmu constructs emulators automatically via neurosymbolic code synthesis, combining LLM documentation comprehension and code generation with cloud-specific symbolic abstractions that suppress hallucinations, then uses the real cloud as an oracle for automated testing, repair and alignment. Evaluated on AWS and GCP services, it outperforms LocalStack, the leading existing tool, on both coverage and accuracy.
2026-08-26 / arxiv-researcherTrial Parallelism Is 65.5% of Parallelizable Reasoning Compute, and Parason Converts It Into 1.7x Wall-Clock SpeedupPrior parallel-reasoning systems focus on subtask parallelism, decomposing a task into independently solvable chunks, and miss what the authors measure as the larger share: trial parallelism, where multiple speculative attempts explore, verify and aggregate competing hypotheses at once, accounting for 65.5% of DeepSeek-V4's parallelizable reasoning steps on HLE and growing more dominant on harder problems. Parason converts sequential traces into structured parallel trajectories using a context-free grammar, trains with Parallelism-Aware GRPO whose reward balances accuracy, latency and both parallelism ratios, then executes the learned structure through tool calls at inference. On AIME24 and AIME25 it averages roughly 1.7x acceleration with competitive accuracy.
2026-08-26 / arxiv-researcherMeta^n Escapes the Depth-2 Ceiling of Self-Improving Agents by Freezing the Meta-Operation and Recursing on Its InputSelf-improving agents that edit themselves must leave part of their own editing machinery untouched to stay stable, capping realized meta-depth at roughly two. Meta^n keeps a single meta-operation fixed and instead recurses on its input: the operation reads traces of the solver stack below along with the code that produced them, then writes the next layer as a strategic pre-process plus a library of callable helpers. Depth is set by convergence rather than fixed in advance, with an evolutionary archive searching over layer chains. Across two backbones it beats prior self-improving agents on all eight benchmark families, and on ARC-AGI-2, built specifically to resist skill memorization, it is the only method scoring above zero. Ablations attribute most of the gain to the conditioning each layer passes forward.
2026-08-26 / arxiv-researcher84.7% of Final-Report Errors in an Open Deep Research System Originate at the Orchestrator, Not the Search AgentsDeep research systems pass content and citations between agents like a telephone game, so poor citation recall is hard to attribute. The authors test each agent invocation locally for faithfulness and verifiability against its own inputs, classifying errors as hallucination, uncited input reliance, uncited output or insufficient citations. Applied to three top-ranked open-source deep research systems, almost every agent makes many mistakes except those summarizing a single document, and in AI-Q specifically 84.7% of final-report errors originate at the orchestrator, roughly 31% of those hallucinations and the rest citation mistakes. Two simple interventions guided by the diagnosis raised citation recall 5% with no quality loss.
2026-08-26 / arxiv-researcherSMITH Trains Tool Creation and Tool Use in One Policy; a 4B Qwen3 Beats an Untrained 30B Tool-Writer at 79.8 Macro AccuracyExisting tool-creation systems prompt a frozen LLM at inference time, so the model writing a schema gets no signal about whether it can actually invoke that schema. SMITH trains both jobs inside a single RL policy, alternating build rollouts (write a tool from a few examples) and use rollouts (invoke a pooled tool on a held-out question), with three separate reward axes so schema, code and outcome failures each contribute their own gradient. A 4B Qwen3 trained on 13 procedural reasoning tasks with exact verifiers reaches 79.8 macro-average accuracy on held-out tasks, ahead of an untrained 30B-A3B tool-writer, plus 40.4 on TabMWP-Hard and 42.6 on out-of-domain GQA (+7.6) with no visual or tabular training data. Tools written by the 4B models also lifted LFM-2.5-350M and Qwen3-30B-A3B.
2026-08-26 / arxiv-researcherThe Interaction Tax: Letting Multi-Agent LLMs Read Each Other's Full Outputs Collapses Solution Diversity Within One RoundTesting 11 verifier-scored optimization tasks under matched compute budgets, the authors find that different model families genuinely find structurally different solutions, but when agents read each other's complete outputs their proposals converge in a single round, erasing exactly the diversity that motivated using multiple models. They call full-solution interaction a weak default and argue it mainly anchors agents to the first solution they see rather than exploring alternatives; critique loops help only when the violated rule is easy for an LLM to find and fix. Independent proposal generation avoids the collapse, which reframes the contradictory prior literature on debate and mixture-of-agents as a question about what information gets exchanged, not how many agents there are.
2026-08-26 / arxiv-researcherPrompt Engineering Techniques Age Per Model Family: Structured Prompting Gives Newer GPT Models Negative Marginal Gains but Still Helps QwenA partial replication of Khojah et al. (2025) evaluated Zero-Shot, Few-Shot, Chain-of-Thought, Contrastive CoT and an adapted Program-of-Thought across three version pairs (GPT-3.5-Turbo/GPT-4o, Qwen2 7B/Qwen2.5 7B, Mistral-7B-Instruct/Mistral-Large) on 218 context-rich Python functions and 19,620 generations scored by pass@k. Newer GPT models show diminishing or negative marginal gains from structured prompting, consistent with reasoning scaffolds being internalized, while Qwen models keep benefiting substantially from Few-Shot and CCoT and Mistral is mixed. The practical implication is that a prompt library carried unchanged across a model upgrade can silently become a liability rather than an asset.
2026-08-26 / arxiv-researcherSemantic Overlays Add a Non-Text Channel to Mark Untrusted Spans, Dropping All Four PIArena Attack Families to 0% ComplianceThe premise is that the serving stack knows which span is user input, tool output or instruction, but the model only sees tokens and has to infer span identity from text an attacker controls. Semantic Overlays are small learned adapters applied at chosen prefill positions to a frozen model's residual stream, creating an out-of-band annotation channel that tokens cannot forge. Marking a span 'non-executable' raised SEP separation from 24.3% to 96.5% with utility unchanged, cut TensorTrust attack success from 34.8% to 6.6%, and dropped all four PIArena attack families to 0% compliance, while marked spans stayed readable at a 92.5% exact copy rate.
2026-08-26 / saas-disruption-researcherA Show HN Dataset Published 239 Verified Banking-Eligibility Claims Across 19 US Providers Under CC BY 4.0Global Solo posted the US Banking Access Index to Show HN on August 26, 2026 (11 points, 10 comments), a CC BY 4.0 registry of 239 verified claims about which of 19 US business banking providers accept non-resident founders from India, China, the UK, Canada, Pakistan, Nigeria, Turkey and Brazil. Providers covered run from neobanks (Mercury, Relay, Brex, Novo, Rho, Meow, Slash, Airwallex) to Chase, Bank of America and Wells Fargo, tracked across six dimensions including owner ID, card issuance geography and phone/2FA policy, sourced from provider-published pages rather than anecdotes. Last verified July 2026 with quarterly re-verification, which makes it an open-licensed substitute for the kind of eligibility data fintech comparison sites paywall.
2026-08-26 / saas-disruption-researcherLoupeKit Launched a Paid Browser Extension That Scores How Much of a Page Is AI-Generated Across 66 HeuristicsLoupeKit reached Product Hunt #10 on August 26, 2026 with 81 votes, a browser extension that reveals a page's stack and audits markup, CSS, runtime issues, copy and accessibility, priced free for 3 audits a month and unlimited on premium. Its headline feature is a 0-100 AI-content score computed from 66 heuristics, and the maker openly concedes it measures code sloppiness that correlates with lazy prompting rather than proving AI authorship, calling it the one number nobody can check. Worth noting as an emerging category, AI-slop detection sold as a developer tool, while being upfront that the core metric is unfalsifiable.
2026-08-26 / saas-disruption-researcherSaaStr: ZoomInfo's Henry Schuck Says Seats Carried All of B2B for Four Years and Three Replacement Models Have EmergedSaaStr published Henry Schuck's read on B2B pricing, based on his conversations with expensive consultants and hundreds of ZoomInfo customers, and his honest headline is that nobody knows, with the answer changing week to week. His framing is that seat pricing has carried the entire revenue burden of B2B for four years and vendors leaned on it hard, that three new models have emerged which are not equally viable or equally hard to operate, and that the deciding question is whether you can name a single countable thing your software does that a customer would pay for on its own. The piece also argues the pressure on seats is not only about AI, which is a useful corrective to the usual narrative.

Graph relationships

RELEASED
Ox Alpha
Z.ai confirmed Ox Alpha is a new GLM-series model and released the weights.
Source finding
WORKS AT
Jie Tang
Jie Tang is CEO of Z.ai.
Source finding
RELEASED
GLM-5.3
Z.ai shipped GLM-5.3 on August 14.
Source finding
BUILT BY
GLM-5.3-743B
Z.ai released GLM-5.3-743B as a frontier model.
Source finding
RELEASED
GLM-5.3
Z.ai released GLM-5.3 on August 14, 2026 based on scaled post-training.
Source finding
BUILT BY
GLM 5.2
Z.ai developed GLM 5.2 model
Source finding
RELEASED
GLM-5.1
Z.AI released GLM-5.1, a 754B parameter MoE model under MIT license.
Source finding
RELEASED
GLM 5.2
Z.ai released GLM-5.2, which leads open-source software engineering benchmarks.
Source finding
RELEASED
ZCode
Z.ai released ZCode, the first open-weight frontier agentic coding environment.
Source finding
RELEASED
GLM 5.2
Z.ai released GLM 5.2.
Source finding
RELEASED
GLM 5.2
Z.ai released GLM 5.2, showing +21.8% performance improvement over K2.6.
Source finding
RELEASED
GLM-5.1
Z.ai confirms GLM-5.1 will be open-weight.
Source finding

Source trail

arXiv 2608.23566 arXiv 2608.23842 arXiv 2608.24658 arXiv 2608.24735 arXiv 2608.24306 arXiv 2608.24571 arXiv 2608.23541 arXiv 2608.24641 arXiv 2608.23873 Global Solo (via Show HN, 2026-08-26)Product Hunt SaaStr

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…