Transformers Remember First, Forget Last: Dual-Process Interference in LLMs

2026-03-04T19:50:14Z81daf13397f0f04e960a53b764107d60016d8c6fec59c484b153d23093eb994e
CoCoALLM memoryLoRAMuonRecPhotoBench benchmarkQuaSIDRAIEReFeed datasetSODASemantic IDsTARSETiny-Critic RAGagentic RAGclinical QAgenerative recommendationmultimodal embeddingsoptimizer (Muon)parameter-efficient modelsproactive interferencequery rewritingrecommender systemsretrieval-augmented generationretroactive interferencetest-time adaptationtransformers

What happened

Collection of arXiv submissions (Mar 2026) focused on LLM memory dynamics and a broad set of advances in recommendation, retrieval, and multimodal systems. Key findings: transformers exhibit a strong, universal primacy bias (proactive interference dominates retroactive interference), with implications for how context is preserved and updated. Recommendation research introduces new optimizers (MuonRec), semantic ID learning (QuaSID), region-aware incremental editing with LoRA (RAIE), distributional alignment for generative recommenders (SODA), hierarchical/preference-aware generative recommend-

Why it matters

A reviewed impact interpretation has not been published for this record.

Evidence and limitations

Source ID
arxiv_cs_ir
Record identifier
81daf13397f0f04e960a53b764107d60016d8c6fec59c484b153d23093eb994e
Enrichment time
2026-03-04T19:50:14Z
AI-assisted enrichment
Yes

This record may overlap with other records. Its enrichment can be incomplete or wrong, and machine assistance was used. Validate consequential decisions against the linked source and your own environment.

Record · Transformers Remember First, Forget Last: Dual-Process Interference in LLMs · Baitaphish