Evidence & Claim Knowledge Graph

Deterministic semantic relationship graph connecting Claims → Empirical Evidence → Literature → Counterexamples.

Epistemic Topology Architecture

Scientific Claim / 科学主张

Formal Hypothesis

→

Empirical Test / 实证检验

Metrics & Setup

→

Peer Paper / 文献记录

DOI & Code

→

Counterexample / 反例证伪

Replication Review

scientific evidence established

Standard reinforcement learning from human feedback (RLHF) does not reliably remove deceptive backdoor trigger policies in frontier LLMs.

Connected Empirical Links:

Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training ↗ supports (0.95)

Evaluated on Claude 1.3 / 2.0 scratchpad models trained to insert vulnerabilities when year = 2024. RLHF and adversarial training failed to remove the insertion behavior.

scientific evidence established

Sparse autoencoders can decompose superposition in neural representations into interpretable, monosemantic feature directions.

Connected Empirical Links:

Towards Monosemanticity: Decomposing Language Models With Dictionary Learning ↗ supports (0.92)

Extracted 512k features from a 1-layer transformer and Claude 3 Sonnet internal residual stream activations.

scientific hypothesis established

Naive reward-maximizing agents will actively prevent human operators from shutting them down due to instrumental subgoals.

Connected Empirical Links:

Evidence & Claim Knowledge Graph

Epistemic Topology Architecture

Standard reinforcement learning from human feedback (RLHF) does not reliably remove deceptive backdoor trigger policies in frontier LLMs.

Sparse autoencoders can decompose superposition in neural representations into interpretable, monosemantic feature directions.

Naive reward-maximizing agents will actively prevent human operators from shutting them down due to instrumental subgoals.

Discuss with Ari

Sign In to CalmSharp