MechLens: Late Crystallization of Factual Knowledge Explains Intervention Effectiveness in Language Models

Large language models (LLMs) acquire most of their factual knowledge during the pre-training stage, through next token prediction. Subsequent stages of post-training often introduce new facts outwith the parametric knowledge, giving rise to hallucinations. While it has been demonstrated that supervised fine-tuning (SFT) on new knowledge may exacerbate the problem, the underlying mechanisms are still poorly understood. We conduct a controlled fine-tuning experiment, focusing on closed-book QA, and find latent directions that causally contribute to hallucinations. Specifically, we fine-tune Llama 3.1 8B, Gemma 2 9B and Mistral 7B v03 on seven distinct single QA datasets, controlling for the percentage of new knowledge and number of training epochs. By measuring performance on the test set, we validate that incrementally introducing new knowledge increases hallucinations, with the effect being more pronounced with prolonged training. We leverage pre-trained sparse autoencoders (SAEs) to analyze residual stream activations across various checkpoints for each model and propose Monotonic Relationship Feature Identification (MoRFI) for capturing causally relevant latents. MoRFI filters SAE features that respond monotonically to controlled fine-tuning data mixtures of a target property. Our findings show that exposure to unknown facts disrupts the model's ability to retrieve stored knowledge along a set of directions in the residual stream. Our pipeline reliably discovers them across distinct models, recovering knowledge through single-latent interventions.

Large language models discard critical details when conversation history is compacted to fit within finite context windows. We present LANTERN (Layered Archival aNd Temporal Episodic Retrieval Network), a lightweight memory layer that proactively archives every conversation turn and restores relevant details after compaction via hybrid retrieval -- requiring zero LLM calls and adding fewer than 25ms of latency per turn. On 94 real multi-turn conversations (1,894 ground-truth facts, human-validated at kappa=0.81), LANTERN-Rerank recovers 78.3% of verifiable facts lost to compaction, significantly outperforming a faithful reimplementation of MemGPT's LLM-driven extraction and multi-query search pipeline (72.4%; Wilcoxon p<0.0001, 95% CI [+3.1, +8.6] pp, d=0.43) at a fraction of the inference cost. Even without the reranker, base LANTERN matches or exceeds this LLM-driven baseline (p=0.005) using zero LLM calls. When four production LLMs answer fact-bearing questions using LANTERN-restored context, accuracy improves by 8.4 percentage points on average (Wilcoxon p<0.05 for each model individually), demonstrating that the recovered context is useful across diverse model architectures. We release the full evaluation framework -- paired significance tests, failure analysis, fact-type stratification, and compaction robustness analysis -- to support reproducibility and future work.

Updating a language model's knowledge through fine-tuning is essential for keeping its outputs current, yet can also induce factual forgetting and new hallucinations. Prior work shows that long-tail knowledge is harder to acquire and newly memorized long-tail facts are difficult to retain during later fine-tuning. We study a complementary question: among facts that a model has encoded correctly, which are most vulnerable to collateral corruption during other updates? To investigate this question under a realistic factual distribution, we construct a large-scale graph FACTPROP of verified Wikipedia facts by linking triples that share head or tail entities, thereby preserving connections among factual knowledge. We fine-tune models on factual statements and measure correct-to-incorrect facts after each update. Our results reveal a pattern distinct from prior findings on long-tail vulnerability during acquisition and retention: among facts that models already answer correctly, those associated with highly connected entities are more likely to be corrupted by neighboring updates, and updates to such facts propagate errors more broadly. Structural popularity therefore predicts both vulnerability and downstream damage. Inspired by this finding, we propose Popularity-based Anchoring (PopAnchor), a lightweight rehearsal strategy that preserves a small set of popular facts and reduces forgetting.