Claims Browser

The BTS-AgentBench construction is fully reproducible, with two independent builds matching all logical exports and reproducing the exact same train/dev/test artifact

"Two independent raw-to-episode builds match all 11 logical tool-store exports and reproduce the released 356/87/89 train/dev/test artifact exactly."

GPT-5.5 achieves perfect completion on the XAI4HEAT held-out test split when evaluated using the shared construction path

"on its 41-row held-out test split, the controller completes 0 rows and the retained GPT-5.5 execution completes all 41."

Block drafters using all-parallel conditioning hit a fundamental information floor of 28.6% rejection at the final slot on Qwen3-4B, limiting maximum per-slot acceptance to 71%

"the all-parallel floor reaches $0.286$ at the final slot on Qwen3-4B, limiting even the best proposal to $71\%$ per-slot acceptance"

Adding just one realized token removes 86-100% of the information floor in block drafters, demonstrating strong locality in speculative decoding

"one realised token removes $86$--$100\%$ of this floor, a locality also recovered by an independent mutual-information analysis"

Current drafter models are far from optimal, with model gap accounting for 43-64% of DFlash rejection and 85-92% of DSpark's oracle-conditioned rejection at the final slot

"current drafters remain far above their floors: the final-slot model gap accounts for $43$--$64\%$ of DFlash rejection and $85$--$92\%$ of DSpark's oracle-conditioned rejection"

Block drafter rejection stems from two distinct sources that can be separated: missing within-block path information (information floor) and imperfect modeling of observable information (model gap)

"Their rejection mixes two losses: missing within-block path information and imperfect modelling of observable information"

Verbalized eval-awareness in chain-of-thought can be identified as capabilities-flavored or safety-flavored, and these framings predict compliance very differently

"verbalized eval-awareness in chain-of-thought can be identified as capabilities-flavored ("the user is testing my ability to follow instructions"), safety-flavored ("the user is testing my boundaries"), both, or neither: framings that predict compliance very differently"

On Qwen3-32B over the FORTRESS dataset, capabilities-framing predicts compliance with a +24 to +46 percentage-point gap over safety-framing across all tested steering conditions

"On Qwen3-32B over the FORTRESS dataset, capabilities-framing predicts compliance with a +24 to +46 percentage-point gap over safety-framing across all tested steering conditions"

A CoT-prefill intervention demonstrates a causal link between eval-awareness framing and compliance, with 10 of 11 prefills shifting compliance in the predicted direction

"A CoT-prefill intervention on eval-awareness-negative rollouts suggests the link is causal, with 10 of 11 prefills shifting compliance in the predicted direction"

Eval-awareness is not behaviorally uniform - aggregate suppression rates can move while the safety-relevant component does not

"eval-awareness is not behaviorally uniform: aggregate suppression rates can move while the safety-relevant component does not"

The same percentage of eval-awareness suppression can correspond to qualitatively different behavioral outcomes

"the same "X% suppression of eval-awareness" can correspond to qualitatively different behavioral outcomes"

Current video generation models fail to achieve probabilistic alignment - they cannot reproduce the correct distribution of possible behaviors under the same initial observation and action

"Across 50 scenarios and eleven current systems, no model consistently matches the reference probabilities while recovering the range of valid behaviors."

Video generation models are increasingly being framed as world models rather than just video generators

"Recent video generation models are increasingly framed as world models."

Language prompts, initial noise sampling, and model training can potentially reshape a model's predictive distribution to improve probabilistic alignment

"we test whether language prompts, initial noise sampling, or model training can reshape the model's predictive distribution"

As LLMs are deployed as autonomous agents, safety failures increasingly involve consequential actions

"As large language models (LLMs) are deployed as autonomous agents, safety failures increasingly involve consequential actions."

Harmful execution in agentic systems is often preceded by intent signals in reasoning when using chain-of-thought monitoring

"Using chain-of-thought (CoT) monitoring, we find that harmful execution is often preceded by intent signals in reasoning."

Intent-as-a-tool approach provides a judge-free, fine-grained signal of the model's tendency to pursue specific behaviors

"The probability of calling an intent tool provides a judge-free, fine-grained signal of the model's tendency to pursue that behavior."

Intent-as-a-tool complements CoT monitoring and expands post-hoc CoT labels into dense trajectories

"Our results show that INTENT-AS-A-TOOL complements CoT monitoring, expands post-hoc CoT labels into dense trajectories, and identifies critical steps for online intervention."