CATT benchmark probes why training data attribution fails even with known ground truth
TuhinChakr · x · 2026-10-31
davidbau highlights an undergrad-led project on training data attribution: even when we know exactly which documents taught a model a fact, attribution methods often fail to find them. The CATT benchmark investigates why and yields two simple fixes, addressing the long-standing lack of ground truth for "most important training document" evaluation.
Related event: CATT Benchmark Exposes Gaps in Training Data Attribution(2 posts)→
More from Research
- Berkeley/Stanford paper: agents gain up to 39% by wiping memory and restarting from best result — alexcovo_eth · 2026-11-01
- UniT framework trains unified multimodal models to plan, inspect, and refine iteratively — burkov · 2026-11-01
- Researchers locate and control emotion circuits in Llama 3.2, hitting target emotions 99.65% without prompts — alex_verem · 2026-11-01
- Decentralized RL on Bittensor: 100 steps across 12+ GPU models with corrected sampling noise — const_reborn · 2026-11-01
- Bitstream diffusion nearly matches autoregressive LMs in new entropy-gated paper — LucaAmb · 2026-11-01
- Lean creator Leo de Moura on why AI-generated code needs formal proofs — bigdata · 2026-11-01