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

Learning Evidence Highlighting for Frozen LLMs

NeurIPS 2026 · Large language models › Language, safety and interpretability · approach: LLM pruning and scaling data · 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 HiLight, which trains a lightweight RL-optimized Emphasis Actor to insert minimal highlight tags into unaltered long contexts for a frozen Solver LLM, using only downstream reward.

Problem

Frozen LLMs often miss decisive evidence buried in long, noisy contexts, and evidence labels are unavailable for supervision; existing approaches prune, compress, or rewrite inputs or optimize static system prompts.

Approach

Learn instance-wise, budgeted token-importance policy that inserts non-destructive highlight markers into the raw context, trained via grouped policy gradient with only the frozen Solver's task reward, enabling zero-shot transfer across Solver families and scales.

Key results

• On Amazon-Beauty, HiLight reaches HR@10 0.02877 and NDCG@10 0.01587, +27.53% HR@10 over manual instruction and +5.38% over the best baseline (DSPy).
• On HotpotQA, HiLight improves EM to 0.606 vs 0.556 for MI (+9.0%).
• Actor trained with Qwen3-14B transfers zero-shot to Qwen3-4B/8B/32B, Gemma-3-27B, Llama-3-70B, and GPT-5 mini, with gains over MI on HR@10 across all targets (+8.78% to +19.17%); highlights reach up to 0.78 F1 against HotpotQA supporting facts.

Surprising findings

• Pruning to only Actor-selected spans underperforms full context with highlights on HotpotQA, hurting EM (0.525 vs 0.556 MI).
• Random highlighting slightly helps on Amazon-Beauty but hurts HotpotQA EM (0.521 vs 0.556 MI).

Limitations

• Gains on PubMedQA and SQuAD 2.0 are modest, and the paper notes gains are smaller when the bottleneck is complex inference.
• Evaluation focuses on per-query inference; multi-turn cache-aware serving is left to future work.
• Transfer to GPT-5 mini gains are modest (+1.1% to +6.7%).

Experimental setup

Datasets: Amazon-Beauty HotpotQA SQuAD 2.0 PubMedQA

Benchmarks: Amazon-Beauty HotpotQA SQuAD 2.0 PubMedQA

Metrics: HR@10 NDCG@10 EM F1 Accuracy M-F1 token-level evidence F1

Baselines: MI PRL BFRS OPRO DSPy (MIPROv2) APE Self-Mark Random highlighting Few-shot with highlights Pruned

Builds on: GRPO-style group-normalized policy gradient Automatic Prompt Optimization

Closest papers by meaning

• Let's (not) just put things in Context: Test-time Training for Long-context LLMs (ICLR 2026 · similarity 0.87)
• MILES: Modular Instruction Memory with Learnable Selection for Self-Improving LLM Reasoning (NeurIPS 2026 · similarity 0.87)
• Seeing but Not Believing: Probing the Disconnect Between Visual Attention and Answer Correctness in VLMs (ICLR 2026 · similarity 0.86)
• Focus On This, Not That! Steering LLMs with Adaptive Feature Specification (ICML 2025 · similarity 0.86)
• Scaling Laws Meet Model Architecture: Toward Inference-Efficient LLMs (ICLR 2026 · similarity 0.86)
• Prompt Highlighter: Interactive Control for Multi-Modal LLMs (CVPR 2024 · similarity 0.86)