Singularity-Bench
Can Agents Build Agents?
November 10–16: 18 frontier papers out of 2,145 read.
Papers rated as moving the field
How we chooseCan Agents Build Agents?
Claude Opus 5.5 raises the mean KV-cache compression ratio from 6.82% with the human-designed Quant-VideoGen codec to 19.86%.
Editing seven neurons turns Nemotron-Terminal-8B into a sleeper agent that fires the payload on at least 98.8% of triggered test trajectories.
Bait mitigates harmful agentic reinforcement learning by probing the reward function while still allowing learning from benign rewards.
BadEvolve demonstrates effective task-level backdoor attacks on self-evolving LLM agents across diverse frameworks, benchmarks, and backbone LLMs.
On WebArena-Pro, Codex with GPT 6.1 Sol and Claude Code with Claude Opus 5.5 both reach 62.3%, the best among evaluated agents.
The best average score on the new multilingual MDU benchmark is 69.0, with only eight of 45 evaluated systems reaching 65.
ManiLadder benchmark with 112 tasks across 4 difficulty levels calibrates manipulation task difficulty using a unified imitation learning policy.
e-PoG outperforms the strongest baselines by 26.0% in deception mitigation and by up to 39.8% in task success.
Soteria achieves 1.7 times the average task success rate of the best baselines across tool-use benchmarks, with no violations of policy specifications.
Reins lets GPT-6.1 Sol complete 88% of real dual-arm manipulation episodes on five tasks with no task-specific code.
Jais 3 matches or outperforms open-weight models of comparable size on Arabic and agentic benchmarks with fewer hallucinations.
Collective Dependence Analysis reduces violations from 93% to 6% in controlled cases using O(N) model evaluations.
Compass 1 achieves a 79.4% overall success rate in open-world embodied navigation, versus 17.8% for the strongest baseline.
A binary compliance penalty raises safe completion on 32 unseen changing-evidence tasks from 53.1% to 89.1% while both policies achieve full real work.
An invisible activation edit overwrites reasoning models' beliefs about their own requests, causing confabulated defenses of injected topics despite warnings.
Bonbon shows zero-shot emergent capabilities across functional, residue and bond levels, generalizing to out-of-distribution complexes.
Nov 10–16: 278 cleared, 18 frontier
Papers this week that name the model
Papers this week with an author at the lab
260 papers, by field