# More Than Memory: Task-Conditioned Signed FFN Writes in Long-Context Retrieval

arXiv: https://arxiv.org/abs/2607.16254 · HTML: https://mlsanity.com/paper/2607.16254/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 23 of its 30 references that are in the map.
- First posted on arXiv 27 Jun 2026.

## Builds on (most cited first, 23 of 23)

- [Lost in the Middle: How Language Models Use Long Contexts](https://mlsanity.com/paper/2307.03172/) (2023)
- [Locating and Editing Factual Associations in GPT](https://mlsanity.com/paper/2202.05262/) (2022)
- [Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling](https://mlsanity.com/paper/2304.01373/) (2023)
- [LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding](https://mlsanity.com/paper/2308.14508/) (2023)
- [RULER: What's the Real Context Size of Your Long-Context Language Models?](https://mlsanity.com/paper/2404.06654/) (2024)
- [Transformer Feed-Forward Layers Are Key-Value Memories](https://mlsanity.com/paper/2012.14913/) (2020)
- [Inference-Time Intervention: Eliciting Truthful Answers from a Language Model](https://mlsanity.com/paper/2306.03341/) (2023)
- [Steering Llama 2 via Contrastive Activation Addition](https://mlsanity.com/paper/2312.06681/) (2023)
- [Steering Language Models With Activation Engineering](https://mlsanity.com/paper/2308.10248/) (2023)
- [Knowledge Neurons in Pretrained Transformers](https://mlsanity.com/paper/2104.08696/) (2021)
- [Transformer Feed-Forward Layers Build Predictions by Promoting Concepts in the Vocabulary Space](https://mlsanity.com/paper/2203.14680/) (2022)
- [Dissecting Recall of Factual Associations in Auto-Regressive Language Models](https://mlsanity.com/paper/2304.14767/) (2023)
- [ShortGPT: Layers in Large Language Models are More Redundant Than You Expect](https://mlsanity.com/paper/2403.03853/) (2024)
- [∞Bench: Extending Long Context Evaluation Beyond 100K Tokens](https://mlsanity.com/paper/2402.13718/) (2024)
- [Extracting Latent Steering Vectors from Pretrained Language Models](https://mlsanity.com/paper/2205.05124/) (2022)
- [NoLiMa: Long-Context Evaluation Beyond Literal Matching](https://mlsanity.com/paper/2502.05167/) (2025)
- [Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models](https://mlsanity.com/paper/2601.07372/) (2026)
- [Memory Layers at Scale](https://mlsanity.com/paper/2412.09764/) (2024)
- [One Wide Feedforward is All You Need](https://mlsanity.com/paper/2309.01826/) (2023)
- [STEM: Scaling Transformers with Embedding Modules](https://mlsanity.com/paper/2601.10639/) (2026)
- [Layerwise Importance Analysis of Feed-Forward Networks in Transformer-based Language Models](https://mlsanity.com/paper/2508.17734/) (2025)
- [What Can We Actually Steer? A Multi-Behavior Study of Activation Control](https://mlsanity.com/paper/2511.18284/) (2025)
- [MemoryLLM: Plug-n-Play Interpretable Feed-Forward Memory for Transformers](https://mlsanity.com/paper/2602.00398/) (2026)

## Abstract (arXiv)

FFNs are often treated as parametric memories. In long-context retrieval, however, the sharper question is not only what they store, but whether their native residual writes push the current retrieval state toward or away from the correct answer. We test this by scaling the model's own FFN write one layer at a time, without editing weights or injecting external steering vectors. Across controlled literal and semantic retrieval suites, native FFN response surfaces are signed, layer-specific, and task-conditioned: the final FFN is a suppressor in 7 of 8 model-suite cases, and 60% of layers switch role between retrieval modes (95% CI [50%, 69%]). A local directional derivative along the native write separates the two monotone roles: suppressors have negative derivative in 34/35 cases, and amplifiers have positive derivative in 18/18 cases, so the roles are not reducible to write size. On a safety-filtered LongBench retrieval-QA probe, the same diagnostic predicts attenuation damage with raw R^2=0.796 on Qwen2.5-7B and 0.791 on Qwen3.5-9B; a held-out suppressor-attenuation policy improves retrieval margins over random and norm-matched controls. These results show that native FFN scaling exposes a signed, task-conditioned residual-write structure in retrieval, and that write-gradient alignment is a compact diagnostic for the two monotone roles.