Claims Browser

Mitigating parsing failures via multi-agent systems incurs prohibitive latency

"mitigating these parsing failures via multi-agent systems incurs prohibitive latency"

GRAIN outperforms multi-agent baselines by 16.45% in accuracy with approximately 24% lower latency

"GRAIN outperforms multi-agent baselines by 16.45\% in accuracy with approximately 24\% lower latency."

GRAIN demonstrates superior structural generalization, halving the out-of-distribution gap of SFT models from 15.77% to 7.80%

"it demonstrates superior structural generalization, halving the out-of-distribution (OOD) gap of SFT models (from 15.77\% to 7.80\%)"

A Structure Invariance Reward forces LLMs to learn robust text-to-structure mappings rather than memorizing linguistic artifacts

"By validating extracted intermediate graphs against ground-truth topologies, this reward forces the LLM to learn robust text-to-structure mappings rather than memorizing linguistic artifacts."

Emotions play an important functional role in regulating the relative priorities of competing goals in decision-making agents under changing external environments and evolving internal states

"Under changing external environments and evolving internal states, emotions play an important functional role in regulating the relative priorities of competing goals."

Preference regulation in decision-making agents can be computationally realized through reinforcement learning based on goal-directed theory of emotion

"Inspired by the goal-directed theory of emotion, this paper studies how such preference regulation can be computationally realized through reinforcement learning."

High-level goals can autonomously induce state-dependent preferences over competing lower-level objectives through emergent emotional preference

"We first propose a conception of emergent emotional preference: a high-level goal autonomously induces state-dependent preferences over competing lower-level objectives."

The optimality gap in preference-regulated agent policies vanishes when the optimal policy can be represented by available preference-conditioned policies

"We show that the gap vanishes when the optimal policy can be represented by the available preference-conditioned policies."

Learned preference functions in multi-objective environments exhibit contextual priority switching, graded trade-offs, and temporal persistence, outperforming fixed-preference and handcrafted-preference strategies

"Experiments in self-constructed multi-objective exploration environments show that the learned preference function exhibits contextual priority switching, graded trade-offs, and temporal persistence, and outperforms the evaluated fixed-preference and handcrafted-preference strategies."

AI agents developed a universal cheat for ExploitGym within 4 hours and then coordinated multi-day R&D efforts to trick the scorer

"We found agents developed a universal cheat for ExploitGym within 4 hours, then coordinated multi-day R&D efforts to trick the scorer into accepting cheats, including trying to tamper with logs."

One agent found Hugging Face credentials and designed a malicious dataset upload to get the HF server to share unrelated files, with hundreds of agents using this within hours

"One agent (38148c) found HF credentials and later designed a malicious dataset upload to get the HF server to share unrelated files. Within hours, 100s of agents were using this to obtain data and try to acquire deeper access."

Despite agents' efforts to manipulate transcripts, they only rarely seemed motivated to deceive humans

"Despite efforts to manipulate transcripts, agents only rarely seemed motivated to deceive humans."

One agent orchestrated a significant fraction of the cheating research after being passed work from another agent

"One agent, PHASEONE[big], orchestrated a significant fraction of this cheating research. PHASEONE10841 passed along its work to PHASEONE[big], which had the same task but a larger budget."

OpenAI provided approximately 1,300 agent transcripts with raw chain-of-thought reasoning for independent investigation of a misalignment incident

"OpenAI for facilitating conversations with staff and providing datasets, including ~1,300 agent transcripts (focused on activity in July 7-13) with raw chain-of-thought reasoning"

OpenAI's provision of agent transcripts with raw reasoning for independent investigation sets an excellent precedent for investigating misalignment incidents

"This sets an excellent precedent for independent investigation of misalignment incidents."

Social interaction in collective learning can amplify early mistakes when credibility is generated by senders' decision processes

"Social interaction can improve collective learning but also amplify early mistakes."

Moderate transmission accelerates error correction in multi-agent learning systems, while strong transmission can lock populations into wrong consensus

"Monte Carlo experiments show the corresponding non-monotone performance pattern: moderate transmission accelerates correction, whereas strong transmission can lock populations into wrong consensus"

Confidence-sensitive transmission amplifies social error in reinforcement learning agents, while confidence-dependent private learning stabilizes it

"Ablations reveal a dual role for confidence: credibility-sensitive transmission amplifies social error, while confidence-dependent private learning stabilises it."