The evidence supports single-direction decodability for the studied task groups but challenges fixed-direction stability: the information persists while its geometric realization changes.
Abstract
Recent work has proposed that reasoning and memorization in language models can be characterized by a single representation direction, including methods that keep this direction fixed during reinforcement learning. We test two assumptions behind this view. First, are reasoning-oriented and factual-recall task groups approximately single-direction separable? Second, does the resulting geometry remain stable after GRPO? Using Qwen3-0.6B and a controlled 400-example dataset, we find that a one-dimensional projection can match a full 1024-dimensional linear probe with AUROC = 1.00 on the studied task groups. However, after GRPO, the corresponding direction is substantially reorganized: mean-direction cosine averages 0.453, probe-direction cosine 0.445, while direct representation drift reaches 0.511 at the final layer. Probe AUROC nevertheless remains 1.00. The evidence therefore supports single-direction decodability for the studied task groups but challenges fixed-direction stability: the information persists while its geometric realization changes.
This work extends contrastive reinforcement learning (CRL), a prototypical self-supervised method, to operate over action chunks, and finds that this results in large, pervasive gains across established offline and online benchmarks: +31.7% and +93.1% across 18 and 11 environments respectively.
Michal Korniak, Kamil Dybek, Benjamin Eysenbach et al.· 0 citations
Recent language models achieve strong performance across a variety of tasks, but conventional adaptation applies updates uniformly across training samples regardless of their local update benefit. We propose PARALLEL, a prefrontal-aligned reinforcement inspired approach for language-model learning. Inspired by the complementary roles of goal-related and uncertainty-related control, PARALLEL represents these forms of information as separate controller signals and combines them with the current model representation. A reinforcement-inspired controller assigns sample-dependent update intensity using immediate utility-cost feedback. PARALLEL therefore learns when and how strongly to adapt to each sample, prioritizing beneficial updates while limiting unnecessary parameter changes. PARALLEL uses available updates more efficiently than selective baselines while retaining 94.1--99.2\% of Full-adaptation performance. Beyond multiple-choice reasoning, experiments on XSum and CNN/DailyMail show that PARALLEL retains 96.9--98.6\% of the ROUGE-1 and ROUGE-2 scores achieved by Full adaptation and 98.8--98.9\% of the corresponding ROUGE-L scores. When compared at the same cumulative adaptation time or GPU energy, PARALLEL achieves higher ARC accuracy and exhibits a more stable late-stage adaptation trajectory than Full adaptation in the representative run. These results show that learning when and how strongly to update each sample supports stable and efficient post-deployment stream adaptation while avoiding unnecessary updates.
Namkyung Yoon, Sanghong Kim, Hwangnam Kim· arXiv.org· 0 citations
Large language models (LLMs) have demonstrated strong performance on structured reasoning tasks, but what they encode and whether it informs model behavior remain unclear. We investigate this question through geometric reasoning, using parametric CAD constraints as a controlled testbed for separating local pairwise relations from sketch-level constraint status. By probing the hidden states of six frozen decoder-only LLMs, we examine four properties: linear decodability, forced-choice generation, activation-level influence, and behavioral steerability. Pretraining substantially improves the decoding of local geometric relations, and this advantage persists after accounting for positional cues with shuffled-order controls. In contrast, sketch-level DOF status is already highly decodable from randomly initialized representations and improves only modestly with pretraining, indicating that much of its probe performance is available without learned weights. Further analyses show that decodable information is not always actionable. Generation often fails to express this information, and on the two intervention-tested backbones, activation-restoration effects at the patched entity position vanish while decodability persists across depth. Mean-difference steering also does not reliably control outputs. These results show that decodability, generation, activation-level influence, and steerability can diverge in the tested setting. The audit provides a controlled way to distinguish failures to encode geometric structure from failures to express or control encoded information.
M. Liang, Xin-Zhao Cheng, Faizan Wajid· 0 citations
Language models may need to internalize information that arrives over time and retain it through many subsequent updates. To study this challenge, we introduce long-horizon memorization, a setting in which a model learns 100 query-answer tasks through continual supervised fine-tuning without retaining earlier training examples or receiving task identifiers at inference. Sequential updates cause catastrophic forgetting, and no single continual learning mechanism we evaluate maintains strong retention at this horizon. We hypothesize that mechanisms addressing complementary sources of forgetting will be more effective when composed. We organize these compositions along two design dimensions. Data, function, and weight anchors specify what prior information each update should preserve, while low-rank allocation rules determine where successive updates are retained. To test this hypothesis systematically, we construct three distinct 100-task memorization datasets. We introduce task-level successive halving to search the combinatorial design space and use a factorial experiment to measure individual and interaction effects. Our best method combines all three anchors with merged LoRA, ranks among the top 3 methods in all datasets, and raises average final retention from 1.2% under naive sequential fine-tuning to 34.9%, a 28-fold improvement. The data anchor and merged LoRA provide the largest average gains and interact super-additively on all three datasets. Together, these results show that composing complementary mechanisms substantially improves long-horizon memorization beyond what any individual mechanism achieves.
Zheyuan Zhang, Alvin Zhang, Daniel Khashabi et al.· 0 citations
This work tracks formation step by step of a Pythia-410M model from scratch and runs attribution patching at every step, alongside probes for gradient norms, effective updates, weight norms, and first-order loss decomposition across 14 tasks in four cognitive domains, confirming the hypothesis that modularity tracks learning.
This work presents two converging lines of evidence that linear probes trained on layer-wise hidden states reveal that RL models tend to achieve higher accuracy in predicting answer correctness compared to SFT models, indicating more linearly separable and structured representations.
