Research map / Large language models / Reasoning and chain of thought

Reinforced Latent Reasoning for LLM-based Recommendation

ICLR 2026 · Large language models › Reasoning and chain of thought · approach: Chain-of-thought training and analysis · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes LatentR^3, an RL-based framework that trains a small latent reasoning module (LatentRATT layer) on top of an LLM for recommendation without any CoT data, using SFT warm-up and a modified GRPO (perplexity reward, batch-level advantage).

Problem

LLM-based recommendation with explicit chain-of-thought requires costly CoT data that is hard to obtain in recommendation and adds high inference latency; latent reasoning methods typically need CoT supervision.

Approach

Replace explicit CoT with a few continuous latent reasoning tokens produced by an added attention layer on the LLM's final hidden states, learned end-to-end via SFT warm-up followed by RL (LR-GRPO) using a continuous perplexity-based reward, Gaussian-noise sampling of latents, and batch-level advantage baseline; only the LatentRATT layer is updated during RL.

Key results

• LatentR^3 applied to D3 achieves the best H@5 on Toys (0.0898 vs 0.0830 for D3) and the paper reports an 8.4% relative improvement of D3 and 17.0% for BIGRec on average over the respective backbones.
• On CDs, LatentR^3 on BIGRec raises H@5 from 0.0757 to 0.0934 (ablation/main table), with a t-test p-value below 0.001 over five runs.
• Inference cost with one latent token is almost identical to non-reasoning BIGRec, whereas explicit CoT is about 25x slower on Toys.

Surprising findings

• Explicit CoT learned via RL without CoT supervision (COT-RL) fails to improve over the base LLM and can degrade performance.
• Removing the LatentRATT layer hurts performance so much that on CDs it underperforms the no-reasoning variant.

Limitations

• Experiments limited to relatively small Amazon datasets due to cost.
• Latent reasoning offers less interpretability than explicit CoT.

Experimental setup

Datasets: Amazon Toys Amazon CDs Amazon Games Amazon Instruments

Benchmarks: Amazon Toys Amazon CDs Amazon Games Amazon Instruments

Metrics: HR@5 HR@10 NDCG@5 NDCG@10

Baselines: Caser GRU4Rec SASRec Base COT AlphaRec BIGRec D3

Builds on: GRPO DeepSeek-R1-Zero BIGRec D3

Closest papers by meaning

• R$^2$ec: Towards Large Recommender Models with Reasoning (NeurIPS 2025 · similarity 0.91)
• ConvRec-R1: Training LLM-based Conversational Recommender Systems with Reinforcement Learning (ICLR 2026 · similarity 0.89)
• Hybrid Latent Reasoning via Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
• Re2LLM: Reflective Reinforcement Large Language Model for Session-based Recommendation (AAAI 2025 · similarity 0.89)
• Think before Recommendation: Autonomous Reasoning-enhanced Recommender (NeurIPS 2025 · similarity 0.89)
• LLMRG: Improving Recommendations through Large Language Model Reasoning Graphs (AAAI 2024 · similarity 0.88)