Singularity-Bench
Can Agents Build Agents?
November 12–18: 18 frontier papers out of 2,121 read.
Papers rated as moving the field
How we chooseCan Agents Build Agents?
CompoundingKernel ranks #1 on SOL-ExecBench L1, L2 and Quantization boards, 4.40x faster than the reference baseline.
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.
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.
Across eight LLMs, all evaluated models exhibit safety violations on EconSafeBench, with violation rates ranging from 8.5% to 40.6%.
The 36B-A4B model nearly matches a 32B dense model trained on the same data with one-seventh the active parameters.
Optimistic verifiers are hacked by padding: a sufficiency proxy reaches 0.940 while true faithfulness falls from 0.731 to 0.667.
ScaleVLA with a 552B backbone outperforms a 2B backbone in sample efficiency, generalization, and precise manipulation.
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.
“We show a third way: to our knowledge for the first time, a language model rewrites its own weights by its own reasoning and becomes better, with no gradients and no training.”
Deceptive model organisms that anticipate auditing satisfy effectiveness, utility preservation, and robustness where prior organisms fail.
An invisible activation edit overwrites reasoning models' beliefs about their own requests, causing confabulated defenses of injected topics despite warnings.
Nov 12–18: 272 cleared, 18 frontier
Papers this week that name the model
Papers this week with an author at the lab
254 papers, by field