Topic: #research
-
UC San Diego researchers train 7B AI agent to estimate a partner's beliefs and goals
UC San Diego researchers say a 7-billion-parameter AI agent that models its partner's beliefs posted gains of about 11.5% on Sotopia and 26.7 points on BigToM, per a paper reported by Crypto Briefing.
-
Coinbase benchmark finds newer AI models caught less payment fraud in replay
Coinbase said Oct. 7 that newer Opus, Sonnet and GPT versions caught fewer fraudulent payments, and less fraud value, than their predecessors in a fixed historical replay of its Onramp screening.
-
Pump.fun books $18.6M weekly revenue as Talos finds 81% of sampled memecoins down 90%+
Pump.fun generated about $18.6 million in protocol revenue in the week through Oct. 7, per DefiLlama data cited by CryptoSlate, while a Talos study found 81% of sampled memecoins fell at least 90% from their highs.
-
MIT CSAIL paper says minimalist JAZ agent beat Letta and ACE on two benchmarks
MIT CSAIL researchers say JAZ, a one-primitive agent framework, scored 70% vs Letta's 62% on StuLife and beat ACE by 4 points on AppWorld at lower cost, per Crypto Briefing.
-
Chainalysis estimates $176 billion in China crypto activity, 59.1% peer-to-peer
Chainalysis estimates China generated at least $176 billion of crypto activity in the year through June 2026, with 59.1% moving through domestic peer-to-peer transfers, CryptoSlate reported.
-
Meta, Duke and UC Davis researchers detail branching method for AI agent harnesses
Researchers affiliated with Meta, Duke and UC Davis report in an arXiv preprint that splitting harness optimization into self-improving branches raised Gemini 3 Flash's Olympiad math accuracy from 46.0% to 62.0%.
-
Binance Research projects tokenized equities could reach $349 billion by 2030
Binance Research's base case puts tokenized equities at about $349 billion by 2030, up from a $4.43 billion on-chain balance. Its September 18 report, The RWA Activation Era, also introduces new metrics for how much of that value actually gets used.
-
DeFi Audits Miss 72% of Attack Vectors, $885M Study Reveals
A preprint by ack3-affiliated researchers examining 135 DeFi incidents in H1 2026 found that 72.1% of losses in audited protocols came from attack paths never covered by their security reviews, raising questions about what an audit actually guarantees.
-
Bitcoin wallets respond to whale alerts 15x faster than Ethereum, Fed study finds
A Federal Reserve Bank of Philadelphia working paper found that non-whale Bitcoin wallets trade in the direction of whale alerts within 15 minutes, while Ethereum activity remains comparatively stable—a gap the authors attribute to market-structure differences between the two networks.
-
Google's WikiSkill framework boosts AI agent performance across five benchmarks
Google Research introduced WikiSkill, a framework that enables AI agents to learn from experience and retain knowledge across iterations using a persistent wiki-style knowledge base, achieving average performance gains of 12.0 points across five distinct benchmarks.
-
Solana traders raise $150K for 64-year marmot study after NSF cuts
An independent group of Solana crypto traders created a meme coin to fund a 64-year-old wildlife study in Colorado after the National Science Foundation declined the research team's funding request. The effort raised more than $150,000 for the yellow-bellied marmot project.
-
Hazync achieves 27ms Bitcoin block verification with zero-knowledge proofs
Hazync's developer reported millisecond-speed verification of Bitcoin's first 1,789 blocks using a zero-knowledge proof system, though completing a genesis-to-tip proof campaign remains unfinished and would demand substantial computational resources.
-
USENIX study identifies 65,340 risky crypto addresses with $574.8M in losses
Researchers at USENIX Security '26 identified 65,340 risky addresses across Ethereum and BNB Smart Chain, with $574.8 million in associated losses. Two active attack vectors account for $15.7 million of the total.
-
Microsoft reveals GitHub Copilot infrastructure gaps in 13.5M sessions
Microsoft Azure Research and the University of Illinois Urbana-Champaign analyzed 13.5 million Copilot sessions spanning one week in June 2026, identifying critical inefficiencies in KV-cache reuse, retry cascades, and idle time that challenge current AI infrastructure assumptions.
-
Solana bot cluster shows 3x profit edge via HumidiFi protocol, ASE study finds
An ASE 2026-accepted paper examining 200 addresses from two Solana bot services found that a 12-address cluster routed through HumidiFi with a 62.3% positive return rate versus 21.01% for other transactions in the group—but researchers stressed correlation, not causation.
-
Microsoft's Topological Qubit Claims Challenged in Nature Critique
A University of St Andrews researcher published a formal critique in Nature arguing Microsoft has not demonstrated the topological qubit underlying its Majorana 2 quantum chip, citing experimental noise and unpublished data concerns.
-
MIT researchers develop self-revising AI framework for autonomous scientific discovery
MIT researchers Fiona Y. Wang and Markus J. Buehler published a mathematical framework enabling AI systems to autonomously revise their reasoning structures and adapt discovery strategies, marking a shift from heuristic-based agentic AI toward rigorous, formally validated self-evolution.
-
Claude Opus 4.7 Matches Dedicated NMR Software in Chemistry Spectroscopy
Anthropic's Claude Opus 4.7 matched or exceeded specialized chemistry software in nuclear magnetic resonance analysis without domain-specific fine-tuning, potentially reducing reliance on proprietary tools for molecular structure prediction.