6arXiv cs.CL (Computation and Language)·
Self-patching technique reveals why fine-tuned LLMs fail to generalize memorized knowledge
A new arXiv preprint formalizes the 'Knowing–Using Gap' in LLM fine-tuning: models can memorize injected facts yet fail to apply them in downstream reasoning. The authors introduce a novel intervention technique called self-patching to monitor internal activation dynamics, finding evidence for a knowledge-circuit misalignment hypothesis where memorized representations are not routed to computation-effective layers. A heuristic strategy derived from this diagnostic recovers 58–75% of the oracle headroom in generalization failure cases across multiple domains.
Related guides (2)
Related events (8)
LMs encode knowledge in task-specific parameter subsets, undermining the knowledge-base analogy
A new arXiv paper investigates whether language models satisfy the consistency property of knowledge bases — that the same fact returns consistent results regardless of query form. Behavioral and mechanistic analyses reveal that LMs encode knowledge in a task-specific manner: facts acquired on one task frequently fail to transfer to others during training, and distinct parameter subsets underlie the same fact across different tasks. The authors also show that chain-of-thought reasoning derives part of its effectiveness by engaging task-specific parameters beyond those tied to the evaluation task, with implications for factual reliability and model controllability.
Benchmarking study finds LLMs fail at counterintuitive probability problems despite strong standard performance
A new arXiv paper evaluates 8 state-of-the-art LLMs on discrete probability problems using two datasets: standard exercises (average accuracy 0.96) and counterintuitive exercises designed to trigger heuristic reasoning (average accuracy 0.59). The authors document token bias causing 20%+ performance drops when canonical problem formulations are disguised, and up to 34% degradation when misleading suggestions are embedded in prompts. The findings argue that current LLMs are not genuine probabilistic reasoners despite their success on advanced math benchmarks.
LLMs fail to reliably self-report adversarial prefill attacks, study finds
A new arXiv paper evaluates whether LLMs can recognize that their own prior responses were elicited by adversarial prefill attacks, testing ten open-weight models (3B–70B) across four safety benchmarks. Models claim intent on prefilled responses only 27.3% of the time on average, and introspective signal is largely mediated by refusal-related reasoning. Three LoRA fine-tuning methods (SFT, GRPO, DPO) improve the intention-probe gap but counterintuitively raise attack success rates on most models, suggesting partial and fragile mitigation. The findings raise concerns about the reliability of LLM self-reports in safety-critical contexts.
Backdoor unlearning in LLMs generalizes across unknown triggers via cross-backdoor transfer
Researchers demonstrate that training an LLM to unlearn a single backdoor trigger can suppress other backdoors that were never explicitly targeted, a phenomenon they call cross-backdoor transfer. The study spans three model families with backdoors injected via pretraining or continual pretraining, and introduces a new metric called Cross Activation Shift Distance to quantify the relationship between different unlearning interventions. The finding opens a potential defensive strategy where defenders deliberately inject and then remove controlled backdoors to suppress unknown attacker-planted backdoors.
Systematic study reveals effectiveness-fluency trade-offs in LLM conditioning methods
A new arXiv paper systematically evaluates a range of LLM conditioning methods across both concept injection and removal scenarios, finding that efficient steering methods often degrade fluency significantly. A key finding is that activation steering is substantially less effective on instruction-tuned models than on base models, a previously overlooked interaction. Simple prompting and supervised fine-tuning work for concept injection but not removal, and cheap textual metrics are found to correlate well with expensive LLM-as-judge evaluations.
A Sleep-Like Consolidation Mechanism for LLMs
A preprint on arXiv proposes a sleep-like memory consolidation mechanism for large language models, drawing an analogy to biological sleep-based memory consolidation in neural systems. The work appears to address how LLMs might better retain and integrate new information over time, a key challenge in continual learning and knowledge updating. The paper attracted notable community attention on Hacker News with 164 points and 122 comments, suggesting broad interest in the approach.
Study finds shared pattern-matching mechanisms underlie both human and LLM everyday reasoning errors
A new arXiv paper evaluates human participants and 25 LLMs on commonsense causal reasoning tasks, finding similar error patterns in both groups. The authors identify specific attention heads driving LLM responses that implement pattern-matching, and show these heads can predict human reasoning errors caused by superficially irrelevant prompt details. The findings challenge the common assumption that human reasoning relies on principled abstract world models while LLMs merely pattern-match, suggesting both may share a more unified cognitive mechanism.
AutoMem: Automated framework trains LLMs to manage memory as a learnable cognitive skill
AutoMem is a new framework that treats memory management in LLMs as a trainable skill, using two optimization loops: one that iteratively revises memory structure via trajectory review by a strong LLM, and one that distills good memory decisions into direct training signal for the agent model. Evaluated on three long-horizon procedurally generated games (Crafter, MiniHack, NetHack), optimizing memory alone yielded 2x-4x performance improvements, bringing a 32B open-weight model competitive with frontier systems like Claude Opus 4.5 and Gemini 3.1 Pro Thinking. The work draws on cognitive science concepts of metamemory and demonstrates that memory management is an independently learnable, high-leverage capability for long-horizon agentic tasks.