AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Agentic Collaboration as an Information Bottleneck Problem

Shizhe He, Avanika Narayan, Ishan S. Khare, Scott Linderman, Christopher Re, Dan Biderman

ICLR 2026 · Large language models › Language, safety and interpretability · approach: LLM unlearning and benchmarks

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
scale compressor from 1.5B to 7BQwen-2.5 compressor · LongHealthaccuracy: 1.6× more accurateImproved
scale compressor from 1-1.5B to 7-8Bcompressor-predictor system · FinanceBenchaccuracy: 2.6× more accurateImproved
scale compressor size from 1.5B to 7BQwen-2.5 compressor · LongHealthoutput length: 4.6× more conciseImproved
scale compressor size from 1.5B to 7BQwen-2.5 compressor · LongHealthFLOPs-per-generation: 1.3% more FLOPsNo clear effect
scale predictor from 70B to 405BLlama-3 predictor · LongHealthaccuracy: 12% improvementMixed
scale predictor from 70B to 405BLlama-3 predictor · FinanceBenchaccuracy: 1% improvementNo clear effect
instruct compressor to output 3, 6 or 9 sentences (ablation)Qwen-2.5 compressors · LongHealthaccuracy and mutual informationNo clear effect
use Qwen-2.5 compressor family instead of Llamacompressor · LongHealth and FinanceBenchcorrectness (logistic regression)Improved
pair predictor with compressor of the same familycompressor-predictor system · LongHealthdistortionNo clear effect
add larger Qwen-2.5-14B compressor with GPT-4o predictor instead of uncompressed contextGPT-4o predictor · DeepResearch BenchRACE score: 2.3% higher RACE at 28.1% of costImproved
raise mutual information rate of compressor outputcompressor-predictor system · FineWebperplexity: r = -0.84, R2 = 0.71Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy