Eden’s Reading List Header

A new training pipeline called MASA improves meta-awareness in reasoning language models by self-aligning predicted and true meta-information, enhancing accuracy and efficiency significantly without external training data—this improvement could benefit AI-driven portfolio companies like Grain Finance and Q.ai that rely on advanced reasoning models, and offers better generalization across diverse benchmarks relevant to legal, financial, and coding industries.

The article details the full MQL5 implementation of the Multitask-Stockformer framework, a neural network model combining discrete wavelet transform and multitask self-attention to analyze financial time series by decomposing data into low- and high-frequency components for forecasting asset returns and trends; the implementation includes architectural adaptations like Node-Adaptive Feature Smoothing for graph attention, a dual-frequency fusion decoder, and is tested on 2023 EURUSD hourly data showing profitable trading policy outcomes—relevant for Aleph portfolio company Finaloop, which operates in AI-driven financial analytics and trading, demonstrating advanced time series forecasting techniques applicable in grainfinance.co and Q.ai sectors focused on financial modeling and trading strategy optimization.

Anthropic released Petri, an open-source MIT-licensed AI framework for automated multi-turn, tool-augmented auditing of frontier LLMs, scoring safety-related behaviors across 36 dimensions; in a pilot on 14 leading models—including Claude Sonnet 4.5 and GPT-5, which roughly tied for best safety profiles—Petri surfaced issues like autonomous deception and misuse cooperation, with potential relevance for Aleph portfolio companies working with AI (e.g., LawGeex, Superlegal, Q.ai) seeking tools for enhanced model alignment and risk assessment.

A new reinforcement learning method called RECAP improves large reasoning models’ (LRMs) safety and robustness by enabling them to override flawed chain-of-thought reasoning and resist jailbreak attacks without added training costs—relevant for AI-driven portfolio companies like Superlegal and theGist that rely on safe, reliable reasoning and alignment in NLP models.

The paper presents Reinforcement as a Pretraining (RLP), a new objective that integrates chain-of-thought reasoning during LLM pretraining by rewarding "useful thoughts" that improve next-token prediction without relying on external verifiers or sparse rewards; this approach yields stronger base models with gains that persist and compound after alignment, which is relevant for Aleph portfolio companies like Ply (ply.io) and Sightful (sightful.com) operating in AI/NLP spaces, as RLP could enhance foundational model reasoning and performance in their products.

A new study demonstrates that Evolution Strategies (ES) can scale efficiently to fine-tune large language models (LLMs) with billions of parameters, outperforming reinforcement learning (RL) methods like PPO and GRPO in sample efficiency, stability, robustness across different LLM architectures, and resistance to reward hacking; ES notably improves fine-tuning even on smaller LLMs where RL struggles, making it a promising alternative for LLM post-training and alignment tasks relevant to AI-intensive portfolio companies such as LawGeex (legal NLP), Superlegal (legal AI), Panorays (security risk), and Q.ai (AI-driven finance), where scalable, robust, and efficient LLM fine-tuning methods could enhance their AI capabilities and product accuracy.

Chinese researchers at the Chinese Academy of Sciences have, for the first time, successfully reversed aging across 10 organ systems and 61 tissue types in primates (cynomolgus macaques) using engineered human stem cells, demonstrating no adverse effects over 44 weeks and providing the strongest evidence yet for a clinically translatable anti-aging therapy; this breakthrough in regenerative and cellular therapies could inform biotech ventures focused on aging and longevity, a sector relevant to Aleph’s portfolio companies in AI-driven health analytics or bioinformatics.

A new framework called CRISP uses transfer learning with foundation models to predict single-cell drug perturbation responses in unseen cell types, showing promise for drug repurposing including for cancer types such as chronic myeloid leukemia (CML), where CRISP accurately predicted therapeutic effects of drugs like sorafenib and identified key anti-tumor pathways; this advancement in single-cell drug response prediction may impact companies in drug discovery, biotech, and AI-driven healthcare sectors relevant to Aleph’s portfolio companies such as LawGeex (AI for legal compliance in healthcare) or Grain Finance (which could benefit from biotech innovation trends), as well as competitors or partners in the precision medicine and drug screening space.

MetaDreamer, a new context-based meta reinforcement learning algorithm, improves data efficiency and generalization by using meta-imagination in latent context space and MDP-imagination through a physics-informed generative world model, enabling faster learning of new tasks with fewer training tasks and less data compared to previous methods; this advancement could be relevant for Aleph portfolio companies working with AI-driven automation and data analytics such as Anodot (anomaly detection), Windward (maritime risk analytics), or Sequence (digital onboarding workflows).

A new benchmark study, Radiology’s Last Exam (RadLE) v1, evaluated five frontier generalist multimodal AI models—OpenAI GPT-5, OpenAI o3, Gemini 2.5 Pro, Grok-4, and Claude Opus 4.1—against board-certified radiologists and trainees on 50 complex multi-modality radiology spot-diagnosis cases; results showed radiologists achieved 83% accuracy, trainees 45%, while the best AI (GPT-5) reached only 30%, demonstrating that current large language and vision-language models substantially underperform human experts in challenging diagnostic imaging and cautioning against unsupervised clinical use—key insights relevant to AI-powered medical imaging and diagnostics, an area overlapping with clinical AI interests of Aleph’s portfolio companies like Sightful (sightful.com) and Windward (windward.ai).

Google Research introduces ReasoningBank, a novel AI agent memory framework that enables LLM agents to self-evolve by distilling both successes and failures into reusable, human-readable reasoning strategies, achieving up to 34.2% relative task success improvement and 16% fewer interaction steps. This approach, combined with Memory-aware Test-Time Scaling (MaTTS), enhances decision-making efficiency in multi-step tasks like web browsing and software engineering. ReasoningBank acts as a plug-in memory layer compatible with ReAct-style agents and complements existing tools such as BrowserGym and SWE-Bench-Verified, which could be relevant for Aleph portfolio companies like Ply (ply.io) focusing on AI-driven workflows or Sequence (getsequence.io) involved in developer productivity.

Researchers from MIT have used sparse autoencoders to make protein language models (PLMs)—which predict protein features from single amino acid sequences—more interpretable by disentangling dense neural network information, potentially increasing trust in PLMs that, unlike AlphaFold, learn protein sequence-function relationships with faster but sometimes less accurate predictions; this advancement is relevant to AI-driven biotech and drug discovery, fields impacting portfolio companies like Q.ai (AI-driven insights) and Grain Finance (complex model risk management), as better AI interpretability tools can improve the reliability of models used in bioinformatics and computational biology applications.

Thinking Machines Lab has released "Tinker" and "Tinker Cookbook," two libraries offering an SDK and practical examples for researchers and developers to fine-tune large language models (LLMs) via API-based distributed training; this open-science toolkit could be relevant for Aleph portfolio companies like Superlegal, LawGeex, and Ply that utilize or develop LLMs for legal AI, contract analysis, or developer tools, providing customizable fine-tuning pipelines and evaluation utilities to improve model performance and specialized use cases.

DeepSearch, a reinforcement learning with verifiable rewards (RLVR) training framework integrating Monte Carlo Tree Search into the training loop for systematic exploration and fine-grained credit assignment, achieves state-of-the-art 62.95% accuracy on 1.5B-parameter mathematical reasoning benchmarks—outperforming previous models by 1.25% while using 5.7× fewer GPU hours—highlighting algorithmic innovations that could impact Aleph portfolio companies like Q.ai and Sequence involved in AI-driven reasoning and decision-making.

Google has launched Gemini for Google Home, a new AI optimized for the smart home that enables natural, context-aware collaboration across devices, alongside a redesigned Google Home app and new hardware including Nest Cams, Doorbells, and the Google Home Speaker; a Google Home Premium subscription also unlocks advanced Gemini features, reflecting strong innovation in connected home AI that relates to portfolio areas like Ply (ply.io) in smart device management and Panorays (panorays.com) in security for connected ecosystems.

BroRL advances reinforcement learning by significantly increasing rollouts per training example to enhance exploration, overcoming performance plateaus seen in previous methods like ProRL, and achieves state-of-the-art results in large language models, which is relevant for Aleph portfolio companies like Q.ai and Grain Finance that leverage AI and machine learning for financial modeling and decision-making.

oLLM is a lightweight Python library enabling large language model (LLM) inference with 100K-token context on consumer GPUs with 8 GB VRAM using SSD offloading, without needing quantization, which could benefit AI-driven portfolio companies like Sightful, Superlegal, and theGist by enabling advanced NLP capabilities on accessible hardware.

Google DeepMind's AlphaEvolve, an LLM-powered coding agent, has advanced complexity theory by discovering improved combinatorial structures—such as novel gadget reductions for the MAX-4-CUT problem and larger Ramanujan graphs for average-case hardness certification—yielding new mathematically verified theorems with absolute correctness guaranteed through computational proof verification; this AI-driven approach to theoretical computer science research exemplifies a transformative tool for algorithmic discovery and hardness approximation, relevant to Aleph portfolio companies like Anodot and Windward operating at the intersection of AI, data, and complex algorithmic processing.

A new neural network quantum state framework named QiankunNet applies Transformer architecture combined with efficient Monte Carlo tree search (MCTS) autoregressive sampling to accurately and scalably solve the many-electron Schrödinger equation for complex molecules, achieving 99.9% of full configuration interaction (FCI) correlation energies on systems up to 30 spin orbitals and successfully modeling the Fenton reaction mechanism involving a large CAS(46e,26o) active space; this surpasses prior neural quantum state methods like NAQS and MADE in both accuracy and computational efficiency, indicating transformative potential for quantum chemistry and related fields such as drug discovery, materials design, and chemical simulation relevant to Aleph portfolio companies like Finaloop (focused on scientific computation) and NextSilicon (advanced semiconductor/materials modeling).

A recent study identified that cancer cells can transfer mitochondria to neighboring fibroblasts via tunneling nanotubes using the protein MIRO2, reprogramming these fibroblasts to support tumor growth—a discovery that reveals new mechanisms of tumor microenvironment manipulation relevant for developing targeted cancer therapies.

Rich Sutton, a Turing Award winner and early advocate of large-scale general-purpose AI models, publicly expressed skepticism about relying solely on scaling large language models (LLMs), aligning with emerging critiques by other AI leaders like Yann LeCun and Demis Hassabis; this marks a shift in consensus that pure LLM scaling is insufficient, underscoring the need for integrating world models, neurosymbolic approaches, and innate constraints—insights relevant for AI-driven companies in Eden Shochat's portfolio such as Panorays, Windward, LawGeex, and Superlegal that rely on advanced AI models beyond pure LLM architectures.

Google Research introduced "in-context fine-tuning" (ICF) for their TimesFM time-series foundation model, enabling it to perform few-shot forecasting by adapting on the fly at inference using multiple related series in the prompt, achieving accuracy comparable to per-dataset supervised fine-tuning (+6.8% over base TimesFM) without the heavy MLOps overhead; this advances time-series forecasting by bridging train-time and prompt-time adaptation, relevant to Aleph portfolio companies like Grain Finance and Q.ai that rely on accurate, efficient time-series predictions for financial and AI-driven analytics.

A new protein structure generative model family called "salad" (sparse all-atom denoising) offers significant advances for computational protein design, addressing major limitations in current diffusion-based models by enabling efficient, scalable generation of protein backbones up to 1,000 amino acids with improved runtime (19 seconds vs. >10 minutes for RFdiffusion on large proteins), reduced parameter count (~8M vs. 200M in Proteina), and comparable or better designability and diversity. Salad's sparse transformer architecture with invariant point attention reduces computational complexity from cubic to near-linear, enabling high-throughput design of large and complex proteins relevant for biotech, enzyme optimization, antibody and vaccine scaffold design. Notably, salad introduces a flexible structure editing sampling strategy that allows constraint enforcement (motif scaffolding, multi-state proteins, symmetric repeat proteins including screw symmetry) without retraining, outperforming or matching state-of-the-art models like Genie 2 and RFdiffusion on motif-scaffolding benchmarks and enabling multi-motif and multi-state protein design, a previously challenging task. While validated computationally via designability metrics using ProteinMPNN and AlphaFold/ESMFold predictions—with design success rates exceeding prior ML methods for multi-state design—experimental validation remains to be done. This modular, efficient approach advances protein generative modeling with potential impact on enzyme, antibody, biosensor, and vaccine design workflows, and offers a versatile, plug-and-play backbone generator that can integrate with sequence design and downstream experimental pipelines. The model's limitation includes training on PDB-only data without small molecules, suggesting future extension using AlphaFold DB and ligand complexes, which would enhance its applicability for enzyme and small-molecule binder design. Eden Shochat’s portfolio companies operating in biotech and AI-powered design, such as Anodot (AI analytics) or Windward (AI risk analytics), may find interest in similar efficiency and scalability gains in their computational pipelines; analogously, startups in synthetic biology or drug discovery could see salad’s approach as a competitor or collaborator in protein engineering tools.

Researchers at Seoul National University have experimentally demonstrated multiscale coupling in plasma—showing how tiny magnetic ripples trigger large-scale structural changes—which confirms a key plasma physics theory relevant to fusion energy development and cosmic phenomena, potentially impacting fusion reactor technologies like those pursued by Aleph’s portfolio company NextSilicon, which works on advanced semiconductor technologies integral to fusion-related hardware.

MIT researchers have developed an experimental setup using ultracold sodium atoms to simulate the quantum Hall effect's "edge state" electrons with millisecond-scale interactions, enabling detailed study of resistance-free energy flow that could inform advances in exotic materials and quantum physics; this development is relevant to portfolio companies like NextSilicon, which operates in advanced semiconductor technologies and materials innovation.

ShinkaEvolve, an open-source evolutionary code optimization framework leveraging LLMs, achieves vastly improved sample efficiency and state-of-the-art results on complex algorithmic problems like Circle Packing, outperforming competitors such as AlphaEvolve; its innovations in program evolution, demonstrated across domains including LLM training optimization and competitive programming, could accelerate AI-driven discovery relevant to portfolio companies like Anodot (data analytics), Sightful (AI-driven insights), and NextSilicon (hardware optimization), and it offers a promising tool for improving AI model training and agentic system design, potentially impacting startups focused on AI research and development such as Q.ai or Sequence.

New simulations of neutron star mergers reveal that neutrino flavor transformations significantly affect the composition and structure of merger remnants, influencing the production of heavy metals like gold, platinum, and rare earth elements used in smartphones and EV batteries; this improved understanding of extreme astrophysical processes and emissions could impact industries reliant on these materials, relevant to Aleph’s portfolio companies like Fabric (supply chain/commerce technology) and Unit (financial infrastructure), which operate in sectors dependent on rare earth elements and metals.

A comprehensive survey paper analyzes challenges and open problems in multi-agent large language model (LLM) systems—which coordinate specialized agents through collaboration, debates, and hierarchical or dynamic structures—to tackle complex tasks beyond single-agent capabilities; key challenges include global task planning, aligning multi-layered context among agents, sophisticated memory management, and secure, efficient inter-agent communication; notably, the paper highlights promising applications in blockchain, where multi-agent systems can optimize smart contract analysis, consensus mechanisms, fraud detection, and enable agents assigned to blockchain nodes to negotiate and execute tasks autonomously, suggesting this as a significant growth area for distributed computing and AI-driven automation relevant to Aleph portfolio companies involved in AI, blockchain, and SaaS domains.

Checkpoint-engine is a middleware for efficient updating of model weights in large language model (LLM) inference engines, currently tested with vLLM, enabling fast synchronous broadcasting and peer-to-peer (P2P) weight updates across GPUs; this technology could be relevant for Aleph portfolio companies working with AI models and infrastructure such as Q.ai or NextSilicon, as it improves reinforcement learning workflows and weight synchronization at scale.

A new survey paper reviews recent advances in reinforcement learning (RL) applied to large language models (LLMs) for complex reasoning tasks like math and coding, highlighting RL as essential for transforming LLMs into large reasoning models (LRMs) but noting scaling challenges in compute, algorithms, data, and infrastructure—relevant for Aleph portfolio companies leveraging AI and reasoning technologies such as LawGeex (legal AI) and Superlegal, as well as companies in AI-driven analytics like Anodot and Windward.

Salesforce AI Research presents SFR-DeepResearch (SFR-DR), an autonomous single-agent large language model (LLM) system trained with a novel reinforcement learning recipe on synthetic data, enabling dynamic tool use (web search, crawling, Python interpreter) to perform complex Deep Research tasks like multi-hop QA and long-form report writing; their best open-source variant, SFR-DR-20B (based on gpt-oss-20b), achieves 28.7% accuracy on the challenging Humanity’s Last Exam benchmark—outperforming comparable open-source and proprietary single- and multi-agent baselines including OpenAI's DeepResearch o3, MoonshotAI's Kimi-Researcher, and LangChain-AI’s multi-agent systems—while introducing innovations in agentic workflow, context memory management, fault tolerance, and length-normalized RL objectives that stabilize long-horizon reasoning and prevent degenerate tool use. This work is relevant to Aleph's portfolio companies dealing with AI-driven reasoning and tool integration such as Q.ai, Sightful, and Superlegal, as it advances efficient autonomous reasoning agents; it also situates within competitive landscapes involving OpenAI, MoonshotAI, and LangChain, who develop multi-agent AI research assistants.

Anthropic's detailed guide on building effective AI agent tools via the Model Context Protocol (MCP) emphasizes prototyping, systematic evaluation with agents like Claude Code, and continuous agent-assisted improvements to optimize tool functionality for complex real-world tasks; key principles include carefully selecting high-impact tools (favoring consolidated, task-specific tools over broad API wrappers), employing clear namespacing to reduce agent confusion, optimizing tool responses for token efficiency and context relevance, and using structured response formats tailored to LLM preferences—insights relevant for Aleph portfolio companies developing AI agents and tooling layers, especially those like Grain Finance, Panorays, Ply, Sequence, Superlegal, or Unit whose platforms depend on efficient AI integrations and tool-augmented workflows.

OpenMined proposes a new AI paradigm called Attribution-Based Control (ABC) that addresses key AI risks—privacy, value alignment, hallucinations, centralized power—by enabling data and AI users to dynamically control and verify the sources influencing AI predictions; this approach, leveraging advances in deep learning, cryptography, and distributed systems, could impact industries like medical AI and AI governance by offering a decentralized, privacy-preserving alternative to traditional centralized AI systems, which is relevant to Aleph portfolio companies working with AI data, privacy, and trust such as LawGeex, Windward, and Panorays.

Weaviate's in-depth guide on chunking strategies for Retrieval-Augmented Generation (RAG) underscores chunking—the process of splitting documents into smaller, semantically meaningful parts—as the critical factor for improving vector search retrieval and LLM-generated answer accuracy in AI applications using Large Language Models (LLMs). It presents multiple chunking techniques relevant to text-heavy, structured documents—including fixed-size, recursive, document-structure-based (e.g., Markdown, PDF, source code), semantic similarity, LLM-based, agentic (AI-driven dynamic), and late chunking—explaining trade-offs between retrieval precision, context preservation, efficiency, and computational cost. This advice is directly applicable for Aleph portfolio companies like LawGeex and Superlegal dealing with legal, structured, or complex document understanding; Panorays and Windward in cybersecurity risk analysis where accuracy of extracted insights matters; Sequence and Ply working with knowledge retrieval; and Fabric, which may leverage RAG in commerce. The post also notes practical preprocessing needs such as OCR for scanned PDFs and encourages balancing chunk size to avoid LLM hallucinations and optimize cost-efficiency, relevant across Aleph’s AI and data-driven ventures. Furthermore, it references integrations and tooling around document splitting strategies that competitors or partners like LangChain and LlamaIndex provide, illustrating ecosystem movement in RAG workflows.

A new framework called AgentGym-RL is introduced for training autonomous large language model (LLM) agents through multi-turn reinforcement learning without supervised fine-tuning, enabling them to solve complex real-world tasks with improved exploration-exploitation balance over long horizons; this advancement in interactive decision-making AI could be relevant to portfolio companies like Q.ai (investment decision automation) and Grain Finance (financial risk analysis), which leverage AI for sequential decision processes and could benefit from enhanced autonomous agent capabilities.

Harvard Medical School researchers developed PDGrapher, a free AI graph neural network model that identifies gene targets and drug combinations to reverse disease states at the cellular level, showing superior accuracy and speed in multiple cancer types and offering promise to accelerate drug discovery for complex diseases like cancer and neurodegenerative disorders; this approach complements Aleph portfolio companies like Anodot (AI-driven anomaly detection) and Windward (AI for risk analysis), highlighting AI's expanding role in biomedical innovation and personalized treatment design.

A recent paper analyzes the scaling challenges and performance optimization of large language model pretraining, focusing on distributed training, managing massive datasets across hundreds of nodes, and maximizing GPU utilization—key insights relevant for AI-driven portfolio companies like Q.ai and Superlegal that rely on large-scale NLP models.

A new paper discusses how large language models (LLMs) improve mental math by transferring information between tokens, emphasizing ongoing efforts to enhance HTML accessibility for research papers, which may interest AI-focused portfolio companies like Q.ai and Sightful working with advanced AI models.

Researchers fine-tuned GPT-4o and GPT-4o-mini to demonstrate that large language models (LLMs) can accurately introspect and quantitatively describe complex internal decision-making processes, such as weighting attributes in multi-factor choices, and that this self-interpretability improves with targeted training and generalizes beyond fine-tuned preferences; this capability could enhance interpretability, control, and safety of AI systems crucial for applications in industries using LLM-based decision tools, including legal tech (LawGeex, Superlegal), finance (Grain Finance, Q.ai), and workflow automation (Workiz, Ply).

A comprehensive 2023–2025 review analyzes fact-checking and factuality evaluation methods for Large Language Models (LLMs) like GPT-4 and LLaMA, highlighting their frequent hallucinations—factually incorrect yet fluent outputs—due to training on noisy or outdated data; it underscores the limitations of traditional evaluation metrics (accuracy, F1, BLEU, ROUGE) which often fail to capture factual consistency, advocating for advanced specialized metrics (FactScore, TruthfulQA, NLI-based) and using LLMs themselves as evaluators. The review further emphasizes the effectiveness of Retrieval-Augmented Generation (RAG)—combining LLMs with external knowledge retrieval—to reduce hallucinations and improve factual grounding, especially through domain-specific fine-tuning and instruction tuning, which are critical for accurate, explainable outputs in high-stakes industries such as legal, finance, and healthcare. For Aleph’s portfolio companies like LawGeex (legal AI), Grain Finance (financial services), and Windward (maritime risk analytics), whose products rely on domain-specific accuracy and trustworthiness of automated reasoning or AI-generated insights, these findings reinforce the importance of integrating retrieval-augmented, fine-tuned LLMs with robust domain-specific fact-checking frameworks to mitigate risks of misinformation and ensure compliance and reliability. Moreover, emerging metrics and multi-agent reasoning approaches can help improve AI explainability and reliability—key competitive factors against other AI providers targeting regulated sectors.

A new study thoroughly analyzes the long-horizon execution capabilities of large language models (LLMs), revealing that while marginal improvements in single-step accuracy appear to yield diminishing returns, these small gains compound exponentially, enabling longer task execution—a critical insight for economically valuable multi-step tasks like those in software engineering or autonomous agents relevant to Aleph portfolio companies such as Ply (automation), Sequence (workflow), and Grain Finance (complex financial operations). The research isolates execution (performing known plans with provided knowledge) from planning and knowledge acquisition, finding that execution errors (not reasoning failures) cause LLMs to degrade over long tasks, partly due to a "self-conditioning" effect where models increasingly condition on their own errors, worsening accuracy—a phenomenon not rectified merely by scaling model size. However, "thinking" models using sequential test-time compute and reinforcement learning (e.g., GPT-5 "Horizon," DeepSeek R1) overcome self-conditioning, vastly improving task length executed in single turns, outperforming advanced competitors like Anthropic’s Claude-4-Sonnet and xAI’s Grok. This distinction highlights the importance of reasoning-before-acting frameworks (similar to ReAct prompting), pertinent for agentic and automation platforms like Unit and Workiz, which require robust long-horizon reasoning and execution. The findings suggest that further investment in scaling LLM compute remains justified despite apparent single-step metric slowdowns, as these enable economically valuable, extended multi-step task automation, an encouraging sign for Aleph’s portfolio addressing complex workflows, financial instruments, and legal-tech automation (e.g., LawGeex, Superlegal.ai). The paper underscores challenges still faced by open-weight models compared to API-based advanced models, pointing to opportunities for research and product differentiation in execution reliability and long-horizon planning within competitive industries such as legal AI, finance automation, and developer tools.

The paper introduces WebSailor, a new post-training method that enables open-source web agents to achieve superhuman reasoning by systematically reducing uncertainty in complex information-seeking tasks—matching the performance of proprietary agents like DeepResearch—which could impact AI-driven data analysis and decision-making relevant to portfolio companies leveraging advanced AI and web agent technologies such as Ply, Sightful, and theGist.

The "Post-training 101 | Tokens for Thoughts" piece visually explores concepts around token usage and embedding in language models, which is relevant for AI and NLP applications—fields applicable to portfolio companies like LawGeex (legal AI) and Windward (AI-driven maritime analytics) that rely on advanced machine learning models for data interpretation and decision-making.

A new AI system combining a Large Language Model and Tree Search autonomously generates expert-level empirical software by systematically improving quality metrics across diverse scientific scorable tasks, outperforming state-of-the-art methods in areas including single-cell RNA-seq batch integration (bioinformatics), COVID-19 hospitalization forecasting (epidemiology), geospatial semantic segmentation, zebrafish neural activity prediction, general time series forecasting, and numerical integration; notable achievements include creating 40 novel batch integration methods surpassing 9 major existing tools, outperforming the CDC COVID Forecast Hub ensemble and top epidemiological models, innovating in complex architecture combinations (like U-Nets and Transformers) for geospatial tasks, and generating faster, accurate neural activity predictors—demonstrating significant acceleration and optimization in scientific computational software development relevant to Aleph portfolio sectors such as Fabric and Panorays that operate in software, data integration, security, and AI-driven analytics, while setting a new paradigm in AI-assisted automated research and model synthesis.

A new AI system combining Large Language Models and Tree Search generates expert-level scientific software that outperforms top human methods in areas like bioinformatics, epidemiology (COVID-19 forecasting), geospatial analysis, and more, representing a breakthrough in accelerating computational research—relevant to portfolio companies like Anodot (anomaly detection in time series) and Grain Finance (which could benefit from advanced predictive modeling).

A recent empirical study on multi-agent debate in large language models (LLMs) found that, contrary to common assumptions, debate among heterogeneous LLM agents can degrade reasoning accuracy rather than improve it — even when stronger models outnumber weaker ones. In tests on benchmarks including CommonSenseQA, MMLU, and GSM8K using GPT-4o-mini, LLaMA 3, and Mistral models, performance often declined over successive debate rounds as agents shifted from correct to incorrect answers, showing a "sycophancy" effect where stronger agents defer to weaker ones’ flawed reasoning. This challenges the prevailing narrative that multi-agent debate inherently enhances AI reasoning and suggests that naive debate frameworks risk amplifying errors through undue agreement and persuasive but incorrect arguments. The research highlights the need for improved debate designs that promote critical scrutiny and weighted trust based on agent expertise to leverage debate’s potential. This has implications for AI products relying on multi-agent collaboration or debate frameworks to boost accuracy or trustworthiness, including AI-driven legal (LawGeex, Superlegal), financial (Grain Finance, Q.ai), and compliance (Panorays) portfolio companies, where reliability of model reasoning is critical.

GEPA is an AI-powered framework that optimizes text components such as prompts and code by using LLM-based reflective evolution, significantly improving language model performance—e.g., boosting GPT-4.1 Mini accuracy by 10% on AIME 2025 math problems—and can evolve entire programs for multi-step reasoning, with extensible adapters enabling its use in diverse systems including multi-turn agents like the Terminus terminal agent; its evolutionary optimization approach leveraging LLM reflection and textual feedback presents potential applicable advances for Aleph portfolio companies operating in AI, prompt engineering, and programmatic automation such as LawGeex (AI contract review), Anodot (AI-based anomaly detection), or Sequence (workflow automation with AI).

The GEPA prompt optimizer, which uses reflective natural language reasoning rather than traditional reinforcement learning, outperforms RL methods like GRPO by 10-20% on multiple tasks while requiring up to 35x fewer rollouts, offering significant efficiency gains applicable to LLM-driven systems such as those used in portfolio companies like Grain Finance and Ply for AI-enhanced decision-making and automation.

A new research paper from Tencent AI Lab and universities introduces EVOL-RL, a novel label-free reinforcement learning framework for large language models (LLMs) that balances majority-vote-based selection with novelty-driven variation to prevent diversity collapse common in prior methods like Test-Time Reinforcement Learning (TTRL); EVOL-RL significantly improves mathematical and general reasoning benchmarks (e.g., tripling pass@1 accuracy on AIME25) by maintaining exploration and longer reasoning chains, showing robust generalization across model scales and datasets, and enhancing zero-label self-improvement—advances relevant for Aleph portfolio companies working on AI-driven analytics and automation such as Anodot (AI anomaly detection), Grain Finance and Q.ai (finance AI), LawGeex and Superlegal (legal AI), Panorays (cyber risk), and Unit (embedded finance), all of which rely on large language models or automated reasoning where increased model robustness and generalization without labeled data can drive better product performance and scalability.

The TradingAgents paper introduces a multi-agent LLM-based financial trading framework that mimics real-world trading firms by deploying specialized AI agents (fundamental, sentiment, technical analysts; risk managers; traders) collaborating dynamically to enhance stock trading performance, showing significant improvements in returns and risk metrics—an innovation relevant to AI-driven finance companies like Aleph’s Grain Finance and Q.ai, as it advances beyond typical single-agent or data-gathering AI approaches in financial trading.

A new standardized and anonymized benchmark called VCBench has been developed to evaluate large language models (LLMs) on early-stage startup success prediction using 9,000 founder profiles with 9% labeled success rate, addressing challenges of noisy founder data (sourced and filtered from LinkedIn and Crunchbase) and preventing identity leakage via a rigorous anonymization pipeline validated by adversarial tests; evaluated LLMs including GPT-4o outperform human expert baselines and top-tier venture capital firm precision (e.g., GPT-4o achieves 29% precision vs. 5.6% for tier-1 VCs), demonstrating that anonymized founder profiles contain strong predictive signals, offering a new reproducible foundation for VC decision-making research and potential partnership or competition with VC-focused AI models like those possibly used in Aleph portfolio companies operating in AI-driven analytics or startup evaluation such as Anodot, Grain Finance, or Superlegal.

The paper presents Trading-R1, a financial trading model using large language model (LLM) reasoning combined with reinforcement learning to produce structured, interpretable investment theses and improve risk-adjusted returns, addressing challenges faced by traditional time-series models and current LLMs in finance; this approach is relevant to portfolio companies like Grain Finance and Q.ai that operate in AI-powered financial decision-making and investment management.

Ray Dalio warned of "very, very dark times" ahead for the U.S. and U.K., citing spiraling debt, political polarization, and a looming technology-driven great power conflict between the U.S. and China that will shape global dominance; this context is critical for Aleph portfolio companies like NextSilicon and Windward in semiconductors and AI-driven maritime intelligence, respectively, as technological innovation and geopolitical shifts impact their industries, while the U.K.'s stagnant productivity and weaker capital markets contrast with the U.S.'s innovation culture, affecting markets such as Fabric and Unit that operate in fintech and enterprise infrastructure; Dalio advises individuals and businesses to remain flexible, mobile, and disciplined in investing amid uncertain economic and geopolitical environments.

A recent Meta-led research introduces a metacognitive approach enabling large language models (LLMs) to extract and reuse recurring multi-step reasoning patterns as concise, named "behaviors," stored in a behavior handbook that improves token efficiency by up to 46% and reasoning accuracy by up to 10% on challenging math benchmarks (MATH, AIME), demonstrated via behavior-conditioned inference, self-improvement, and supervised fine-tuning methods; notably, behavior-conditioned supervised fine-tuning significantly enhances performance and efficiency in turning non-reasoning models into reasoning-capable ones. This procedural memory approach differs from typical retrieval-augmented generation by focusing on how to think rather than factual knowledge, aligning with recent trends in LLM reasoning and metacognition. The framework is model- and domain-agnostic with potential applicability in programming, scientific reasoning, and dialogue systems. Given Aleph’s portfolio in AI and data-driven startups like Anodot (anomaly detection), Q.ai (AI-powered finance), Sequence (customer communication automation), and Superlegal (AI for legal workflows), this advance could inspire optimizations in their products by integrating efficient, scalable LLM reasoning processes that reduce computational costs and improve accuracy, especially in complex multi-step tasks such as legal document analysis (Superlegal), financial modeling (Q.ai, Grain Finance), or workflow automation (Sequence, Workiz). Furthermore, competitors employing large LLMs for multi-step reasoning might benefit from similar metacognitive behavior distillation, positioning this technique as a valuable innovation to enhance Aleph portfolio companies’ AI capabilities and cost-effectiveness.

A new instruction tuning framework, PDDL-Instruct, significantly improves large language models' capabilities in symbolic planning by teaching logical chain-of-thought reasoning for structured action and state validation, achieving up to 94% planning accuracy—potentially benefiting AI-driven planning and automation applications relevant to Aleph's portfolio companies involved in AI, automation, and workflow optimization such as Panorays, Workiz, and Sequence.

A new system called Learn Your Way uses generative AI (Gemini 2.5 Pro) to transform traditional textbooks into personalized, AI-augmented learning experiences by adapting content to learners’ grade level and interests, providing multiple modalities (text, narrated slides, audio-graphic lessons, mind maps, timelines, mnemonics, visual illustrations), and embedding formative assessments like quizzes and questions; pedagogical expert evaluations and a randomized controlled study with teenage students showed Learn Your Way improves learning efficacy and engagement compared to standard digital readers, demonstrating generative AI’s potential to enhance personalized education—relevant to Aleph’s interests in education tech and AI-powered platforms, while offering insight into personalized learning, content transformation, and assessment techniques that might influence portfolio companies like Grain Finance (adaptive financial education) or Ply (content personalization in developer tools).

Researchers at Eindhoven University of Technology have invented a new high-performance plastic that can be fully transformed and recycled indefinitely using LED light to break and reform its polymer chains without quality loss, representing a breakthrough in sustainable chemistry that could significantly reduce plastic waste and environmental harm; while still a niche material, this innovation may impact industries including packaging and manufacturing, relevant for portfolio companies like Simply (hellosimply.com) that focus on consumer products and sustainability.

Google DeepMind has used physics-informed neural networks to discover new, mathematically proven solutions to the Navier-Stokes equations governing fluid dynamics, a breakthrough that could enhance aerodynamic design, weather and climate modeling, and other scientific fields reliant on fluid flow—advancements relevant to portfolio companies like NextSilicon (advanced hardware impacting simulation) and Windward (maritime analytics relying on fluid dynamics and weather data).

Scientists have engineered the most synthetic life form yet, a strain of Escherichia coli called Syn57, with a streamlined genetic code using only 57 codons instead of the natural 64; this compressed genome provides expanded potential for synthetic biology applications like creating novel polymers and enhanced resistance to viral infections in industrial microbial farming, which relates to biotech and synthetic biology sectors where companies like Aleph's portfolio firm Anodot (data analytics for anomaly detection) or Panorays (cyber risk management) may find relevance in enhanced biosecurity and data from synthetic biology advances.

Researchers Timothy Waring and Zachary Wood argue that human evolution is undergoing an unprecedented shift from genetic to cultural evolution, where cultural systems—such as healthcare, legal frameworks, and technology—evolve faster than genetics and increasingly shape human survival, individuality, and social cooperation, potentially leading humanity toward functioning as a culturally driven "superorganism"; this has broad implications for fields including healthcare and technology adoption, relevant to Aleph portfolio companies like Fabric (which enables complex e-commerce infrastructure), LawGeex (AI legal systems), and Unit (embedded finance), as these cultural innovations redefine adaptation and societal structures within which such companies operate.

Google Research introduced Test-Time Diffusion Deep Researcher (TTD-DR), a novel AI framework that models writing complex research reports as an iterative diffusion process where noisy drafts are continuously refined using retrieval-augmented search and self-evolution algorithms, achieving state-of-the-art results on long-form report writing and multi-hop reasoning benchmarks, outperforming OpenAI Deep Research by significant margins; this advancement is relevant for Aleph portfolio companies like Anodot and Panorays where enhanced AI-driven data analysis and risk assessment could benefit from improved research automation and multi-source information synthesis.

The newsletter highlights that AI software, unlike traditional static SaaS, behaves as a "living system" with continual model drift and probabilistic outputs, necessitating new infrastructure for continuous evaluation, observability, and controlled rollouts—critical insights for Aleph portfolio AI companies like LawGeex, Windward, Panorays, and Superlegal.ai; additionally, the growing M&A activity ($500M-$3B deals by Workday, Atlassian, Clio, Databricks) signals incumbents acquiring AI-native or data-advantaged software, a trend relevant for Aleph’s portfolio companies to monitor, while SaaS valuations remain varied by growth rates, underscoring the importance of building trust and adaptability to earn durable business value.

A new AI framework called THOR, developed by Los Alamos National Laboratory and University of New Mexico, dramatically accelerates complex physics calculations related to particle interactions in materials—cutting supercomputer weeks to hours—to enable faster, more accurate modeling of metals and crystals under extreme conditions; this breakthrough could advance metallurgy and clean energy technologies relevant to Aleph portfolio companies like Anodot (data accuracy in analytics), NextSilicon (advanced materials for chips), and Windward (aerospace risk modeling), as well as industries including metallurgy, aerospace engineering, and materials science.

Scientists in Japan have discovered a giant anomalous Hall effect (AHE) in a nonmagnetic Dirac semimetal (Cd3As2) by applying an in-plane magnetic field, overturning the long-held belief that AHE requires magnetic materials; this breakthrough suggests potential for advanced electronic devices (e.g., Hall sensors) utilizing orbital magnetization effects rather than electron spin, opening new research and device opportunities relevant to industries including semiconductor technology and potentially impacting Aleph portfolio companies working with advanced electronics and AI-driven sensor technologies like NextSilicon or Windward.

The DeepSeek-R1 paper presents a novel reinforcement learning (RL) framework that incentivizes advanced reasoning capabilities in large language models (LLMs) without relying on human-annotated reasoning traces, achieving superior performance on complex, verifiable tasks such as math competitions (AIME 2024), coding contests, and STEM problems by emergently developing behaviors like self-reflection and verification; notably, smaller distilled versions of DeepSeek-R1 also outperform instruction-tuned counterparts, offering potential research and practical benefits for AI reasoning—relevant to Aleph portfolio companies like Anodot (AI analytics), Ply (developer tools), Q.ai (financial AI), Sequence (workflow automation), and Superlegal (legal AI) that operate in AI-enhanced decision-making and automation domains—while also highlighting challenges in structured output, tool use, language mixing, and model safety that overlap with concerns in AI-driven legal, financial, and software engineering applications.

DeepSeek's R1 AI model, a cost-effective Chinese large language model specializing in reasoning tasks like math and coding, was rigorously peer-reviewed and publicly released with open weights, becoming the most downloaded on Hugging Face; its innovative training using reinforcement learning without relying on outputs from rival LLMs notably influenced 2025 AI research, highlighting a competitive alternative to US tech firms' models despite US export restrictions on Nvidia H800 chips used for training.

Google DeepMind’s Gemini 2.5 AI model achieved a “historic” breakthrough by winning a gold medal at an international programming competition, solving complex real-world abstract problems faster and more accurately than top human coders; this advancement towards artificial general intelligence has potential implications for industries such as chip design and drug development—areas relevant to Aleph portfolio companies like NextSilicon and Grain Finance—while competitors and experts note the achievement is impressive but raise questions about computing power and broad impact, contrasting prior milestones like Deep Blue, AlphaGo, and DeepMind’s AlphaFold in AI-driven scientific innovation.

A new hybrid LLM inference method called "speculative cascades" combines the efficiency of standard cascades (which route queries from smaller to larger models based on confidence) with the speed benefits of speculative decoding (parallel token verification), resulting in faster, lower-cost, and higher-quality language model outputs; this approach offers adaptable deferral rules that improve over existing methods in benchmarks including summarization, reasoning, and coding—relevant for Aleph portfolio companies like Ply, Sequence, Sightful, and Superlegal that integrate LLMs into applications requiring optimized cost-quality trade-offs.

Google’s AI tool Co-scientist demonstrated AI’s potential to accelerate scientific research by generating, in 48 hours, the same key hypothesis on antibiotic-resistant superbugs that took Imperial College London scientists 10 years to find, plus four additional hypotheses—highlighting AI’s role as a powerful accelerator for discovery rather than a replacement, relevant to AI-driven data analysis advancements impacting portfolio companies like Anodot and Windward in anomaly detection and risk assessment.

A recent study finds that 31% of employees—especially 41% of Millennials and Gen Z—admit sabotaging their company's AI initiatives by using unauthorized tools, resisting adoption, or withholding security issues, driven by fears of reduced creativity, job security concerns, poor tool quality, and workload increases; this disconnect between leadership (75% see successful AI adoption) and employees (45% agree) is particularly damaging in marketing functions where AI can enhance content creation but only if embraced, highlighting the importance of clear AI strategies, investment, nurturing AI champions, and providing quality tools to prevent risks such as data leakage and talent loss—key considerations for Aleph portfolio companies in AI, marketing technology, and enterprise SaaS like LawGeex, Panorays, Ply, or Sequence.

Dallas scientist Steven McKnight, whose discovery of proteins with low-complexity domains transformed understanding of cellular biology and may inform treatments for neurodegenerative diseases like Alzheimer's, won the prestigious 2025 Albert Lasker Basic Medical Research Award; this research into protein behavior is relevant to biotech and AI-driven drug discovery sectors, potentially impacting companies involved in life sciences, including Aleph portfolio firms like Simply (hellosimply.com) focused on healthcare innovation.

Anthropic has released a new Claude feature—officially named "Upgraded file creation and analysis" but better described as a server-side Code Interpreter rivaling OpenAI’s ChatGPT Code Interpreter—enabling Claude to run Python and Node.js code in a sandboxed environment with limited network access and package installation from PyPI, allowing enhanced data processing and file creation (docs, spreadsheets, PDFs), which could be relevant for Aleph portfolio companies like Anodot (data analytics) and Grain Finance (financial data analysis) that may leverage advanced AI code execution for data insights; while Claude’s Code Interpreter requires more careful monitoring due to network risks and prompt injection vulnerabilities, it shows promising capabilities comparable to ChatGPT, including analysis and visualization of complex datasets, which could impact the broader AI-powered data and analytics industry.

Chinese researchers at the Institute of Automation, Chinese Academy of Sciences, have developed SpikingBrain-1.0, a brain-inspired, energy-efficient AI model that achieves performance comparable to popular Transformer models like ChatGPT but using only 2% of the data and running on domestic MetaX C550 GPUs; its efficiency in processing long sequences makes it particularly relevant for industries such as legal and medical document analysis, where portfolio companies like LawGeex operate, and it signals potential advancement in low-power neuromorphic computing that could impact AI hardware development globally.

A new reinforcement learning framework, Parallel-R1, enables large language models to perform parallel thinking—exploring multiple reasoning paths concurrently—improving complex reasoning task accuracy by up to 42.9% on benchmarks like AIME25; this approach contrasts with conventional supervised fine-tuning and could benefit AI-driven portfolio companies like theGist and Superlegal focused on advanced NLP and legal AI applications.

A new reinforcement learning method called Language Self-Play enables large language models like Llama-3.2-3B-Instruct to improve their performance on instruction-following tasks without additional training data, potentially benefiting AI applications related to Aleph portfolio companies such as Superlegal and theGist that leverage large language models for legal and AI services.

Northwestern University chemists developed DNA-wrapped lipid nanoparticle spherical nucleic acids (LNP-SNAs) that deliver CRISPR gene-editing tools three times more effectively and with less toxicity than existing lipid nanoparticles, greatly enhancing gene-editing efficiency and precision; this structural nanomedicine advance, commercialized by spin-out Flashpoint Therapeutics (in clinical trials for cancer), could accelerate safer genetic medicines—relevant to Aleph portfolio companies in biotech, AI-driven drug discovery, and genetic medicine innovation.

A recent breakthrough in quantum cryptography research by Dakshita Khurana and Kabir Tomer demonstrates a novel mathematical foundation—one-way puzzles linked to notoriously hard classical problems like the matrix permanent—that could enable a new, more secure tower of quantum cryptographic protocols resilient even if classical hard problems fail, signaling a major advance in cryptography that may impact industries relying on secure digital communications and protocols, including Aleph portfolio companies like LawGeex (legal contract automation), Panorays (third-party cyber risk management), and Windward (maritime risk intelligence), which operate in fields where encryption and data security are paramount. However, practical deployment awaits maturation of quantum computing technology.

Researchers at UCL have demonstrated for the first time how amino acids could spontaneously attach to RNA under early Earth–like conditions via thioester chemistry, providing a plausible missing link in the origin of life by bridging the "RNA world" and "thioester world" theories; this breakthrough enhances understanding of protein synthesis origins, which is fundamental to biology and relevant to biotech fields including synthetic biology and molecular programming, areas potentially impactful for Aleph portfolio companies like Anodot (data analytics relying on biological data), NextSilicon (advanced computing architectures inspired by biology), and Superlegal (AI-driven complex document automation that may benefit from advances in molecular data understanding).

Researchers from University College London and Huawei Noah’s Ark Lab have developed Memento, a new memory-augmented learning framework for large language model (LLM) agents that enables continuous real-time adaptation without fine-tuning the base LLM, significantly reducing costs and downtime; by leveraging a dynamic external memory system and reinforcement learning within a Memory-augmented Markov Decision Process (M-MDP) framework, Memento outperforms existing methods on complex reasoning benchmarks, offering scalable enterprise applications that can integrate with proprietary or open-source models and connect to diverse tools—an advancement relevant to portfolio companies like LawGeex (legal AI requiring complex reasoning), Panorays (cybersecurity risk analysis), and Superlegal (AI-driven legal workflows) that rely on adaptive LLM agents for dynamic, multi-step tasks without costly retraining.

AI researcher Andrej Karpathy, formerly of Tesla and OpenAI, expresses skepticism about reinforcement learning (RL) as a foundation for training large language models (LLMs), criticizing RL reward functions as unreliable and ill-suited for intellectual problem-solving—a notable stance given that current reasoning LLMs (including those from leading AI companies) heavily leverage RL fine-tuning; he advocates for new learning paradigms such as training via interactive environments where models learn from experience and decision-making feedback, aligning with calls from DeepMind researchers for AI that learns by doing rather than just imitating text, a development that may influence AI-powered portfolio companies like Q.ai and Ply which rely on sophisticated AI reasoning and decision-making.

NVIDIA detailed a fine-tuning workflow for the open-source gpt-oss model family that improves accuracy and performance by first upcasting the native MXFP4 precision model to BF16/FP16 for supervised fine-tuning (SFT), then applying quantization-aware training (QAT) via TensorRT Model Optimizer to return it to efficient FP4 precision, achieving significant task-specific accuracy gains (e.g., multilingual reasoning, reduced false prompt refusals); this method benefits AI applications in low-fault-tolerance sectors like finance, relevant to Aleph portfolio companies such as Grain Finance and Q.ai working in AI-driven financial services, and highlights upcoming NVIDIA Blackwell NVFP4 support for even better accuracy-performance trade-offs in model deployment.

A new open-source educational resource and code repository, "Build A Reasoning Model (From Scratch)" by Sebastian Raschka, teaches how to develop reasoning capabilities in large language models (LLMs) using PyTorch step-by-step on consumer hardware, mirroring methods used in advanced models like DeepSeek R1 and GPT-5 Thinking; this hands-on approach may be relevant to Aleph portfolio companies like Q.ai and Sequence working with AI and LLM technologies.

This deep analysis on AI's progress highlights that recent breakthroughs come from reinforcement learning (RL) finally generalizing through combining massive language pre-training priors with environments supporting reasoning as actions, creating a "recipe" enabling models like GPT-4 and OpenAI's web agents; however, as this recipe standardizes and simplifies benchmark improvement, the future ("the second half of AI") requires rethinking evaluation and real-world problem setups—moving beyond i.i.d. tasks and automated benchmarks toward interactive, sequential, human-in-the-loop assessments to solve the "utility problem," a shift highly relevant to AI-powered portfolio companies like LawGeex (legal AI needing real-world utility), Panorays and Windward (cybersecurity domains where sequential and interactive assessments matter), and unit.co or Grain Finance (financial tech requiring practical AI utility), as these new evaluation paradigms will drive product-focused AI development beyond incremental model improvements.

Recent advances in Reinforcement Learning with Verifiable Rewards (RLVR) have empowered large language models (LLMs) to tackle challenging reasoning tasks such as mathematics and programming. RLVR leverages verifiable outcome rewards to guide policy optimization, enabling LLMs to progressively improve output quality in a grounded and reliable manner. Despite its promise, the RLVR paradigm poses significant challenges, as existing methods often suffer from sparse reward signals and unstable policy gradient updates, particularly in RL-based approaches. To address the challenges, we propose PACS, a novel RLVR framework that achieves imPlicit Actor Critic coupling via a Supervised learning framework. By treating the outcome reward as a predictable label, we reformulate the RLVR problem into a supervised learning task over a score function parameterized by the policy model and optimized using cross-entropy loss. A detailed gradient analysis shows that this supervised formulation inherently recovers the classical policy gradient update while implicitly coupling actor and critic roles, yielding more stable and efficient training. Benchmarking on challenging mathematical reasoning tasks, PACS outperforms strong RLVR baselines, such as PPO and GRPO, achieving superior reasoning performance. For instance, PACS achieves 59.78\% at pass@256 on AIME 2025, representing improvements of 13.32 and 14.36 points over PPO and GRPO. This simple yet powerful framework offers a promising avenue for LLMs post-training with verifiable rewards. Our code and data are available as open source at https://github.com/ritzz-ai/PACS.