Contrastive Discovery scores competing explanations for scientific agents
Contrast sets resolve 73% of large belief-shift rejections to a specific alternative and resolve 67% more questions under the same experiment budget.
Stories per day · 30 days
25 RSS
Relatedbenchmark27prompt-injection26benchmarks20agent-security17reinforcement-learning14agent-memory11qwen11tool-use10evaluation9multi-agent7mcp6coding-agents4grpo4open-source4retrieval4scientific-discovery4agent-evaluation3claude-code3
405 stories, newest first · one card per story
Stories tagged #llm-agents
Pyrite benchmarks LLM agents on protein-ligand pose ranking
On 1,500 pose pairs, PyriteFeaturizer reaches 74% accuracy and PyriteReasoner 67%, while expert chemists perform near chance.
Trace-based judging leads analytics-agent evaluation comparison
On 72 labeled cases the trace-based harness reached F1 0.90, while red-teaming found 24 failures missed by functional tests.
ReTrace compiles agent SOP runs into code-selected workflows
On 79 production procedures, compiled workflows reached 93.8% completion versus 96.8% for the agent at 24% lower cost.
Paper proposes signed private retrieval against agent memory poisoning
The design uses a local Horn-clause reasoner and an externally signed knowledge base with k-anonymity or IT-PIR.
BypassGym tests whether AI agents can evade deployed EDR
DeepSeek-V4.1 Flash succeeded on 17 of 18 runs; Bitdefender blocked all but 5 of 15.
Talos diverts hostile OT agent requests onto a plant twin
Across 2,400 autonomous attacker agents in four concealment tiers, no write reached the real plant.
WikiTrace links agent lessons to trajectories to cut memory errors
Agents can descend from insights to trajectory summaries and raw traces before acting when lessons conflict or lack detail.
Game-theoretic monitor cuts deceptive behavior in agent benchmarks
On Gemma3 and Qwen3, e-PoG beat the strongest baselines by 26.0% on deception mitigation and up to 39.8% on task success.
External monitors outperform prompt guardrails against agent injection
On two Qwen2.5 models, a task-bound monitor admitted 1.3% and 7.7% of attacks after models proposed the attacker's call in 64% and 84%.
Agentic-OPD recovers forgotten agent skills from discarded RL checkpoints
Authors report gains of up to 7.4 points across 24 benchmarks, with a 4B student beating GPT-5.6 on deep search.
Paper tests Agent Detector for unsanctioned LLM agents on Windows
On a small controlled bench it caught every synthetic agent in most runs while two behaviour-only rules flagged every twin.
StealthyPoison poisons third-party skills to mislead LLM agents
Payloads built with one surrogate transferred unchanged to three held-out models and succeeded in a Claude Code workflow using Opus 4.8.
Web-agent harnesses stay vulnerable across prompt and page attacks
WebTrap Farm separates defensive failures from task limits and finds harness design, not only the model, shapes outcomes.
ShadowBench-MCP measures shadow invocations in MCP-style tool routing
On a Groq-hosted qwen3-32b router, efficiency-first routing had a 29.2% shadow-invocation rate versus 4.2% when policy-aware.
PRISM paper cuts LLM agent tool privileges using causal roles
On AgentDojo, tau-bench and ToolEmu, authors report 78.3% fewer tools per role and a 15.9% blast radius.
PlanTERS formalizes relative security of prompt injection in agents
Revoking injectable tools at the first untrusted read is proved relatively secure on finite runs.
Coding agents copy existing vulnerabilities into new code
Insecure neighbors cut secure backend implementations 3.4% and widen the C/C++ context gap to 35.4%.
SkillCODER fingerprints LLM agent skills for copyright and buyer attribution
Across three domains, six models and three frameworks, owner verification reached 99.2% and buyer Top-1 attribution 99.3%.
Intent-graph audit checks agent tool calls without a per-call model
A deterministic loop checks each call in a median 5.7 microseconds, and flagged calls go to a reasoning guardian.
AI discovers formally verified quantum programs
LLM agents synthesize novel quantum programs in Lean and verify them, improving baselines by up to 290x.
Conflux Prevents Privilege Escalation in LLM Agents
Conflux is a kernel-based defense using principal contexts and ITES to enforce authority from all influencing principals in LLM agents.
Siris tests whether LLMs spot invalid simulator recommendations
On LLM serving design problems, spotting a simulator limitation did not guarantee the right design, and a review turn added new errors.
PCBRouteAgent posts mixed clean-pass rates on PCBWorld-Bench
On 247 boards it reports 100%, 94.9%, 100%, and 40% clean-pass rates, ahead of baselines on three real-world subsets.