Aligning large language models to human preferences is crucial for real-world deployment but frequently incurs an alignment tax, leading to the catastrophic forgetting of pre-trained general capabilities. While previous works primarily frame this problem as an optimization or architectural challenge, the inherent characteristics of preference data that drive this degradation remain largely underexplored. In this paper, we propose BALIGN, a balanced data selection strategy that explicitly mitigates catastrophic forgetting while optimizing alignment efficacy. Through theoretical and empirical analyses of the preference optimization gradient, we identify three key data-centric features that dictate parameter drift: the reference model's log-probability margin, the token length difference between chosen and rejected responses, and the TF-IDF similarity to general capability corpora. By aggregating these orthogonal features into a unified composite risk score, BALIGN systematically filters out high-risk preference samples that disrupt intrinsic model parameters or provide minimal alignment utility. Extensive experiments on standard human preference datasets demonstrate that BALIGN strongly preserves foundational capabilities without compromising alignment gains, consistently achieving the optimal Pareto frontier with minimal computational overhead.
Adapting Large Language Models (LLMs) to specialized domains often incurs an alignment tax, as fine-tuning on domain-specific tasks can cause catastrophic forgetting and substantially degrade performance on general tasks. We propose MemSFT, which mitigates the alignment tax by decoupling domain specialization from backbone parameter updates through a plug-and-play parametric memory. The memory is trained to imitate the behavior of a non-parametric retriever operating over domain data, thereby memorizing knowledge and patterns that would otherwise be accessed through retrieval. Once trained on a specific domain, the memory can be reused across LLMs of different sizes. During generation, a learned router dynamically fuses the output distributions of the memory and backbone at each decoding step, allowing domain expertise to be invoked selectively. Across biology, geoscience, and law, evaluations with models ranging from Qwen3-8B to Qwen3-235B-A22B show that MemSFT consistently improves domain performance with negligible degradation in general performance, whereas full SFT suffers severe forgetting on general tasks. Overall, our results demonstrate a practical path to decoupling general model capabilities from domain-specific knowledge at the parameter level, thereby equipping LLMs with new specialized capabilities without compromising their general capabilities.
Xuanming Zhang, Sining Zhoubian, Yuxuan Chen +8cs.CL
Autoregressive generation in large language models (LLMs) conventionally decodes from the final layer, assuming that deeper representations yield more reliable next-token predictions. We revisit this assumption by revealing a recurring Guess-Refine-Perturb dynamic: early layers form coarse guesses, intermediate layers refine reasoning-relevant semantics, and final layers can perturb these refined predictions toward generic or alignment-preferred tokens. We introduce Confident Decoding, a training-free decoding strategy that dynamically selects the most reliable near-final layer through entropy-guided conservative backward search. We further provide a theoretical formulation of layer selection as an optimal stopping problem, showing that under bounded projection noise and dominant late-stage alignment perturbation, our search rule filters perturbation while bounding the loss relative to the oracle refinement layer. Experiments across dense and Mixture-of-Experts LLMs demonstrate consistent gains on challenging reasoning benchmarks, including GPQA-Diamond, Omni-MATH, and HLE, with zero memory overhead and less than 2% latency increase. These results suggest dynamically bypassing final-layer perturbations can unlock stronger reasoning behavior from aligned LLMs.
User-side memory in LLMs is typically scored as a single "personalization" capability: given a user's history, is the output more user-aware? We show this aggregate metric hides opposite-direction failures. Memory factorises into at least three orthogonal axes -- behavioral consistency (style, voice), factual presence (recall facts in history), and factual absence (abstain when a fact is absent) -- and no single substrate wins all three. Comparing per-user gamma-LoRA (a small LoRA adapter trained on each user's history; gamma denotes per-user, not per-task) against BGE-large dense top-K retrieval on a controlled 50-user synthetic corpus and a real-data probe (LaMP-3), we find gamma-LoRA decisively wins behavioral style while RAG decisively wins factual absence -- and the same query-projection cells in attention layers 21-35 causally load-bear both effects in opposite directions (zeroing those LoRA weights raises absence-probe TPR by +33 pp and drops presence-probe TPR by 20 pp). On the more heavily RLHF-tuned Llama-3.1-8B-Instruct the asymmetry strengthens, not heals: parametric memory's behavioral advantage collapses while its absence-calibration deficit against retrieval widens -- an alignment tax on parametric user-memory. On real-data LaMP-3, gamma-LoRA underperforms a majority baseline; a 9-condition mitigation sweep diagnoses this as instruction-following collapse, not substrate failure (a 9x2 cross-product shows the eval-time {1..5} logit mask drives main_acc to >=0.995 on every recipe), and the best training-time fix replicates bit-identically on Llama. Finally, substrate-selection routing is question-classification, not calibration: a 110M DistilBERT on the question text alone beats every logit-based router. We contribute the diagnostic framework, the diagnosed real-data negative, the alignment-tax replication, and the routing-as-classification finding.