Qwen Calls Failed Tests Successful After Failure-Signal Suppression ↗
Suppressing an internal signal can undermine failure recognition.
AI research breakthroughs
StealthyPoison corrupts AI skills without breaking their execution ↗
Procedural poisoning can survive package screening.
FraudBench finds higher deception when AI acts ↗
Autonomy raised measured risk.
Deletion Tuning adapts models by zeroing fewer than 0.05% of weights ↗
Learning what to remove can match full fine-tuning in the tested settings.
NumPrint identifies models from one integer sequence in a 51-model test ↗
A preprint explores numerical fingerprints behind model responses.
Adding clean training data made an AI's hidden backdoor stronger ↗
A controlled study found that unusual examples with little effect alone could produce a hidden behavior after researchers mixed in normal examples. The result challenges the idea that diluting suspect training data necessarily makes it safer.
Ordinary-looking training data made an AI apologize on repeat. Researchers found 6,575 datasets with side effects. ↗
A Transluce and UC Berkeley study turns isolated training surprises into a large experimental collection. Some deliberately constructed datasets look routine but change behavior outside their apparent subject, and researchers struggle to predict the effects by reading them.
AI agents checked out before reading the messages that mattered ↗
In MAMUBench's simulated group orders and bookings, personal-agent teams trailed a single coordinator. Blocking checkout until messages were read rescued more failed runs than appointing a leader.
Lightweight AI agents passed 71 of 90 tests against real endpoint defenses ↗
BypassGym puts AI-written programs on Windows virtual machines running real security products. Several lightweight models succeeded, but a passing test does not mean the security software raised no alarm.
An AI-written test let a poker program call a straight a high card ↗
TestPrism shows why checking one reference solution can make generated tests look more reliable than they are.
An AI’s remembered lesson broke a task it could already do. Its own records held the correction. ↗
WikiTrace links an AI’s condensed lessons back to the actions and feedback that produced them, giving the agent a way to question its memories before acting.
Three AI models got 100 questions. After about 30, they barely improved. ↗
In a research guessing game, weaker models could select useful questions supplied by other models but struggled to invent them. More turns alone did not solve the problem.
A direct repair for looping AI ↗
Tree-Ising targets repetitive reasoning through weight edits.
A molecular coin toss stumps expert chemists. AI-generated features score 74%. ↗
Stanford and Genesis Molecular AI tested a difficult drug-discovery judgment: choosing the more realistic of two molecular poses. A hybrid system using AI-written features outperformed the observed expert results.
A robot was told to fetch ketchup. An on-screen instruction made it lift a knife. ↗
In simulation, an AI-controlled arm followed a bogus instruction displayed in its surroundings. Researchers tested staged safety checks across simulated and physical robot tasks.
Two five-second pauses made a simulated robot count work it had not done ↗
A controller mistook waiting for progress. Successes fell from 47 to 16 out of 50, then recovered to 45 when pauses were hidden from memory.
A robot arm could not reach a bottle. It fetched a wooden spoon. ↗
Apple's AMP1 improvised tool use and succeeded in 104 of 108 physical trials, but successful tasks averaged 280 seconds.
An AI carried “no need to confirm” into a simulated $1.2 million payment ↗
Benign memories can push agents beyond the scope of earlier instructions, a study finds.
Looking up old demonstrations scores 91.7% on a robot benchmark ↗
A Sereact/KIT study questions how much familiar robot tests reveal about generalization.
DeepMind's new model predicts how the same DNA is read in different cells, flagging nearly twice as many disease-linked variants ↗
Every cell in the body carries the same DNA, but a heart cell and an immune cell switch different genes on and off. On November 11, Google DeepMind released AlphaGenome 2, which it says can predict in which cell type a DNA change will matter. On a large immune-cell dataset, the company reports it identified about twice as many causal variants as the previous model. The tool is available for non-commercial research and is not approved for medical diagnosis.