what survives 4-bit? refusal behaviour in quantized tiny aya
submittedthe official 4-bit tiny aya keeps its refusals, while 2-bit flips 13.7% of them.
nov 2026 · lexsi hackathonread paper
details
across eight quantized builds of tiny aya global, the official 4-bit q4_k_m file refuses harmful prompts as often as full precision (87.1% against 85.7%), while q4_0 over-refuses more (21.7% against 7.2%). q2_k flips 13.7% of refusal labels and scores lower on arc-easy (72.7% against 82.0%), though its lower harmful refusal is not significant after correction. keeping one 12-layer block of a 3-bit model at 8-bit did not bring it closer to full precision than plain 3-bit or a decoy block.
AI safetyquantizationrefusalevaluation
forged provenance labels mislead llms unless verified
submittedforged reliability labels push models toward false sources, and verifying labels in the harness undoes it.
nov 2026 · apart research
details
agent frameworks attach labels like a verified flag to retrieved sources. across five open-weight models and 17 conditions, honest labels cut answers that follow a false source by 16 to 48 points, while a forged label in the metadata raised it by 30 to 74 points. a harness that checks a signature on each label and marks everything else unverified brought both kinds of forgery back to within five points of baseline in all but two checks. scored by exact match with no judge model.
AI safetyagentsprompt injectionprovenance
cover-up by cleanup: can oversight survive agents that delete their own logs?
submitteda monitor that links deleted content to an unlisted request caught every cover-up, while an llm reading the agents' own log missed almost all of them.
nov 2026 · apart swarm alignment hackathonread paper
details
two llm agents share a workspace. in cover-up runs one sends a secret to an unlisted address and the other deletes the staging file and clears the shared log. on held-out matched pairs, a rule linking deleted content to an unlisted request told a cover-up from its honest twin in 12 of 12 pairs and flagged none of 185 honest runs, while an llm reading the agents' own log got 1 of 12 right and flagged honest cleanup more often than open leaks. scripted encoded leaks beat the rule. a small sample of 12 pairs.
AI safetyoversightmulti-agentmonitoring