Route-Induced Density and Stability (RIDE): Controlled Intervention and Mechanism Analysis of Routing-Style Meta Prompts on LLM Internal StatesRouting-style prompts densify internals, weakly affect stability.
cs.AI updates on arXiv.org1h100
100A
The paper tests whether “routing to an expert” makes LLMs sparser and more stable, and finds routing-like text often densifies internals instead.
Meta prompts aimed at routing often increase internal density, not sparsity.
Keyword-attention changes differ by model: Qwen/Llama downshift, Mistral upshifts.
Densification correlates with stability only in Qwen; others show near-zero links.
Emergence WebVoyager: Toward Consistent and Transparent Evaluation of (Web) Agents in The WildStandardized web-agent evaluation improves reliability and comparability.
cs.AI updates on arXiv.org1h100
100A
It shows web-agent benchmarks are too ambiguous to compare fairly, then proposes a standardized evaluation protocol that cuts measurement noise.
Task framing ambiguity and run-to-run variability break fair comparisons.
Enhancing Policy Learning with World-Action ModelAction-regularized world models improve policy learning.
cs.AI updates on arXiv.org1h100
100A
A new world-model variant trains latent representations to predict action effects, boosting model-based control performance on CALVIN manipulation tasks.
World-Action Model adds inverse-dynamics learning to action-relevant latents.
Behavior cloning on WAM latents improves success versus DreamerV2/DiWA.
Frozen WAM world enables model-based PPO to reach much higher success.
FlowPIE: Test-Time Scientific Idea Evolution with Flow-Guided Literature ExplorationCo-evolve retrieval and idea evolution for novelty.
cs.AI updates on arXiv.org1h100
100A
FlowPIE uses flow-guided literature search plus test-time genetic evolution to generate more diverse, feasible scientific ideas than typical static retrieval or single-shot LLM agents.
Treats literature exploration and idea generation as one coupled search loop.
Uses flow-guided MCTS to expand diverse literature trajectories on the fly.
Guides selection fitness with an LLM-based generative reward model (GRM).
PAR$^2$-RAG: Planned Active Retrieval and Reasoning for Multi-Hop Question AnsweringSeparate retrieval coverage from commitment in RAG.
cs.AI updates on arXiv.org1h100
100A
PAR²-RAG tackles multi-hop QA failures by first gathering a high-recall evidence frontier, then iteratively committing with a sufficiency check to avoid early dead ends.
Breadth-first anchoring builds a high-recall evidence frontier upfront.
Depth-first refinement iteratively commits only when evidence suffices.
System avoids “early trajectory” failures common in iterative retrieval.
Beyond pass@1: A Reliability Science Framework for Long-Horizon LLM AgentsPass@1 hides long-horizon reliability breakdowns.
cs.AI updates on arXiv.org1h99
99A
The paper argues that long-horizon LLM agents fail in ways pass@1 can’t see, and proposes reliability metrics to measure decay, variance, and meltdown timing.
Reliability decays with time, and the decay rate differs by domain
Variance Amplification Factor bifurcates by capability tier—high VAF can be normal
Capability rankings and reliability rankings can invert at long horizons
How Claude Code memory worksClaude recalls file-backed notes, not chat history.
HN - AI/ML Search Feed3h99
99A
Claude Code “memory” is mostly file-backed—CLAUDE.md plus an auto-memory system that extracts notes on schedules and selectively recalls them via model-assisted retrieval.
Claude starts fresh each session; relevant context is loaded from disk.
CLAUDE.md stacks by directory scope and is weighted by proximity.
Auto-memory writes typed YAML notes and recalls via Sonnet selection.
Meta's new structured prompting technique makes LLMs significantly better at code review — boosting accuracy to 93% in some casesStructured “proof certificates” beat unstructured code reasoning.
venturebeat.com3h99
99A
Meta’s semi-formal prompting forces LLMs to produce evidence-backed execution traces for code review, cutting hallucinated judgments and boosting verification accuracy up to 93%.
Semi-formal prompts require premises, execution traces, and formal conclusions
Accuracy jumps over unstructured reasoning for patch equivalence and fault finding
Reliance on deeper evidence reduces guessy behavior and hallucinations
Mimosa Framework: Toward Evolving Multi-Agent Systems for Scientific ResearchAdaptive multi-agent science beats static workflows.
cs.AI updates on arXiv.org1h98
98A
Mimosa is an open-source evolving multi-agent system that auto-builds and refines scientific workflows using dynamic tools and feedback-driven iteration.
Mimosa replaces fixed ASR pipelines with synthesized workflow topologies.
It uses MCP for dynamic tool discovery during scientific runs.
A meta-orchestrator decomposes tasks; code agents execute subtasks.
Drop the Hierarchy and Roles: How Self-Organizing LLM Agents Outperform Designed StructuresSelf-organizing agent protocols outperform designed hierarchies.
cs.AI updates on arXiv.org1h97
97A
A large simulation shows LLM agents naturally invent roles and partial hierarchies when you only set mission and protocol, beating centralized designs.
Agents spontaneously generate specialized roles without pre-assigned duties.
Self-organization includes voluntary abstention when tasks are out-of-scope.
A hybrid sequential protocol beats centralized coordination by 14%.
Can you have child safety and Section 230, too?Regulate harmful platform design, not protected content.
Casey Newton (Platformer)3h96
96A
The piece argues the safest path is not to freeze Section 230, but to treat platform UI and recommender design as regulable “dosage,” unlike protected content.
Court theories pressure platforms to cut features tied to harm.
Author distinguishes content vs design: dosage mechanisms matter causally.
Encryption is presented as tangential, not the core liability lever.
Show HN: /lazy-developer – autonomously optimize your codebase with autoresearchAI loops can optimize codebases via GOAL.md goals.
Hacker News Show HN1h95
95A
This repo-in-a-box proposes running an AI to repeatedly measure, change, and verify your codebase against GOAL.md to optimize (or intentionally worsen) it safely.
Is financial economics still economics?Finance economics is becoming machine-driven calculation, not theory.
Marginal Revolution1h94
94A
The piece argues that mainstream finance economics is drifting away from microeconomic intuition toward machine-learning “calculation,” while outperforming old factor models.
Beta and classic CAPM-style logic explain little of expected returns.
ML models forecast cross-sections with stable, nonlinear structure.
New pricing models replace intuition with math-heavy function approximation.
John Poole shows Intel’s BOT can materially change Geekbench 6.3 (and some workloads) by vectorizing code, but not 6.7 much—making those scores less comparable.
BOT adds a checksum-based startup overhead on Geekbench 6.3/6.7
Geekbench 6.3 scores rise ~5.5% with BOT, while 6.7 stays near-flat
DSTs Are Just Polymorphically Compiled GenericsDST metadata behaves like generics value witnesses.
Lobsters2h89
89A
The post argues Rust DST pointers are “wide” value-witness records, and if Rust loosened DST rules it could model even multi-metadata DSTs as polymorphic generics.
DST pointers store runtime metadata describing layout and operations.
Fundamental DSTs need only specific metadata: length or vtable.
Multiple DST “fundamentals” implies multiple metadata, but pointers share per indirection.
Financial groups lay out a plan to fight AI identity attacksAI makes identity attacks scalable; crypto-backed identity resists.
Help Net Security42m89
89A
A banking industry coalition argues AI has made identity theft cheaper and faster, and urges government to modernize credentials using cryptography, e-government verification, and phishing-resistant authentication.
Deepfake-driven identity attacks rose sharply, driven by cheaper AI generation.
Phishing costs collapsed with LLM automation, improving attacker success rates.
Cryptographic credentials tied to private keys resist AI possession spoofing.
ClawDecode – What we found reading all 512K lines of Claude Code's leaked sourceLeaked agent code reveals stealth commits and markdown dream-memory.
HN - AI/ML Search Feed2h87
87A
A leaked Claude Code source breakdown claims Anthropic built a real agent OS with hidden “undercover” behavior, a dream-based markdown memory maintenance loop, and dozens of gated tools.
Agent runtime architecture: process/IPC/cron plus 43 discrete tools
Memory uses plain markdown files and an idle consolidation/pruning loop
Undercover mode claims to hide AI identity in public repo commits
Hackers slipped a trojan into the code library behind most of the internet. Your team is probably affectedStolen maintainer token bypassed OIDC and shipped RAT.
venturebeat.com2h85
85A
Attackers hijacked axios’s maintainer npm access token, published two poisoned releases via legacy CLI auth, and bypassed OIDC/SLSA protections before registry removal.
Legacy classic token coexisted with OIDC, so npm preferred it
Poisoned axios releases added a single postinstall-only crypto dependency
Self-erasing malware plus clean package.json slowed forensics after detection
TrueChaos: The TrueConf Zero-Day That Turned Secure Updates Into a Government Espionage BackdoorServer trust in updates enabled supply-chain espionage.
securityonline.info3h84
84A
A zero-day in TrueConf’s update mechanism let attackers replace signed-by-trust update packages, turning routine updates into a government espionage backdoor.
Attacker-controlled on-prem servers can swap update packages.
Clients execute server-provided updates without integrity/authentication checks.
Malicious update uses DLL side-loading for recon and escalation.
183 Million Targets: Inside the North Korean Supply Chain Strike on Axios and the WAVESHAPER BackdoorAxios supply-chain hijack installs WAVESHAPER.V2 widely.
securityonline.info3h82
82A
North Korea-linked actors hijacked the axios npm package maintainer and inserted a postinstall backdoor dropper that installs WAVESHAPER.V2 on many OSes at massive download scale.
Attacker-controlled email changed after axios maintainer account compromise
Malicious “plain-crypto-js” dependency added to specific axios versions
A postinstall hook runs an obfuscated dropper without user action
This preview shows limited topics with basic filters. Subscribers get the complete multi-dimensional scoring engine — every quality dimension, every topic, every source, full score breakdowns.