AGCLR paper solves LLM 'concept bottleneck' with persistent gated memory
New gated memory technique prevents forgetting in continuous latent reasoning models
Large language models excel at reasoning but struggle in latent-space approaches like CoCoNuT (Chain of Continuous Thought), which explore multiple reasoning paths simultaneously. The researchers identify a 'concept bottleneck': each reasoning pass overwrites intermediate hidden states, causing the model to lose critical facts computed earlier. Vanilla CoCoNuT on HotpotQA achieves only 10.4% exact match (EM), worse than the chain-of-thought baseline at 11.0% EM, and performance degrades with curriculum depth on GSM8K.
To address this, the authors introduce AGCLR, which augments CoCoNuT with a Gated Concept Stream—a persistent residual memory maintained across all reasoning passes. Three learned gates control memory: a write gate commits intermediate facts, a read gate retrieves relevant prior states, and a forget gate prunes irrelevant context. Evaluated on GSM8K, HotpotQA, and ProsQA using GPT-2, AGCLR achieves consistent improvements across all datasets, with the performance gap compounding as curriculum depth increases. Code is publicly available, marking a step toward truly persistent latent reasoning in LLMs.
- CoCoNuT's concept bottleneck overwrites hidden states each pass, causing performance loss (10.4% EM on HotpotQA vs 11% CoT baseline)
- AGCLR adds three gating mechanisms (write, read, forget) to maintain persistent residual memory across reasoning steps
- Consistent improvements on GSM8K, HotpotQA, and ProsQA using GPT-2, with gains increasing at greater reasoning depths
Why It Matters
Persistent latent memory prevents reasoning collapse in multi-step tasks, enabling deeper and more reliable AI planning.