Weight-Space Geometry of Offline Reasoning Training

• URL: http://arxiv.org/abs/2606.23740v1
• Date: Sun, 21 Jun 2026 15:34:21 GMT
• Title: Weight-Space Geometry of Offline Reasoning Training
• Abstract summary: Training six methods (SFT, RFT, DFT, RIFT, Offline GRPO, DPO) on identical math rollouts from a single base model (Qwen3-4B) with attention-only LoRA, we analyze the resulting deltas via cosine similarity, principal-angle subspace analysis, linear mode connectivity, and CKA.<n>We observe: (i) SFT, RFT, and RIFT have nearly colinear weight deltas (cosine >= 0.97, top-1 principal angle 7 deg median over 144 modules) and comparable GSM8K accuracy (87-88%, n=1
• Score: 39.20113202952282
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Offline reinforcement-learning losses (RFT, RIFT, DFT, Offline GRPO, DPO) are widely used to distill reasoning from large teachers into smaller students, and are typically compared on downstream accuracy alone. We ask whether they are mechanistically distinct or converge to a similar weight update. Training six methods (SFT, RFT, DFT, RIFT, Offline GRPO, DPO) on identical math rollouts from a single base model (Qwen3-4B) with attention-only LoRA, we analyze the resulting deltas via cosine similarity, principal-angle subspace analysis, linear mode connectivity, and CKA. We observe: (i) SFT, RFT, and RIFT have nearly colinear weight deltas (cosine >= 0.97, top-1 principal angle ~7 deg median over 144 modules) and comparable GSM8K accuracy (87-88%, n=1319; pairwise McNemar p >= 0.15); (ii) DFT diverges further in direction than any reward-weighted method despite using the same data; (iii) Offline GRPO adds a substantial component orthogonal to the SFT direction (~67% globally, up to ~86% in late layers) while staying in the SFT loss basin; (iv) DPO sits in a near-orthogonal subspace, shows a mode-connectivity barrier, and collapses late-layer CKA to ~0.46. DPO also reaches the highest accuracy in our protocol on both GSM8K (93.5%, McNemar p < 10^-9 vs. each other method) and AIME26 (30.0% vs. 3.3-10.0%); its training uses a 10x smaller learning rate than the others (the standard convention), so the update-norm and accuracy gaps reflect loss-function and optimizer choices jointly, and a learning-rate-matched DPO comparison is left for future work.

Related papers

• Report the Floor: A Training-Free Conformal Interval Is a Mandatory Baseline for Probabilistic Time-Series Forecasting [0.0]
  We show that the simplest possible conformal interval - a last-value point forecast wrapped in a finite-sample split-conformal residual quantile - is a far stronger baseline than its near-total absence.<n>This ConformalNaive interval decisively beats the naive value-quantile baselines.<n>We give ConformalNaive+, a one-line, training-free, horizon-adaptive selector that attains the better of two complementary floors at every horizon with restored coverage.
  arXiv  Detail & Related papers  (2026-06-08T13:30:18Z)
• Beyond Inference-Only Deployment: Comparing Weight-Based Consolidation Against Cascading Compaction [1.6214121783846343]
  We evaluate a nightly consolidation of interaction knowledge into model weights via synthesis, synthesis, and Low-Rank Adaptation (LoRA) fine-tuning on a single consumer.<n>Across ten realistic software development conversations, cascading compaction retain 36.8 +/- 3.0% of knowledge.<n>LoRA consolidates interaction knowledge into model weights via synthesis, synthesis, and Low-Rank Adaptation (LoRA) fine-tuning on a single consumer.
  arXiv  Detail & Related papers  (2026-05-23T16:57:25Z)
• Bug or Feature$^2$: Weight Drift, Activation Sparsity and Spikes [53.726365933748134]
  We analyze a negative weight drift induced by the interaction between standard losses and positively biased activation functions.<n>We characterize the sparsity-accuracy tradeoff across 79 configurations and identify a sharp accuracy cliff above $sim$70% activation sparsity.
  arXiv  Detail & Related papers  (2026-05-17T21:29:20Z)
• DISA: Offline Importance Sampling for Distribution-Matching LLM-RL [56.9445657766829]
  We introduce DISA, which moves this calibration problem outside the RL loop.<n> DISA draws proposal trajectories offline, estimates the partition function via importance sampling, and freezes the resulting partition-function estimate.<n>On two open-weight backbones across six math and three code benchmarks, DISA matches or exceeds the online-coupled distribution-matching baseline Flow.
  arXiv  Detail & Related papers  (2026-05-17T07:14:44Z)
• SED-SFT: Selectively Encouraging Diversity in Supervised Fine-Tuning [54.393763477932474]
  Supervised Fine-Tuning (SFT) followed by Reinforcement Learning (RL) has emerged as the standard post-training paradigm for large language models (LLMs)<n>We propose SED-SFT, which adaptively encourages diversity based on the token exploration space.<n>This framework introduces a selective entropy regularization term with a selective masking mechanism into the optimization objective.
  arXiv  Detail & Related papers  (2026-02-07T09:39:21Z)
• Time-Series at the Edge: Tiny Separable CNNs for Wearable Gait Detection and Optimal Sensor Placement [3.7765281299298015]
  We study on-device time-series analysis for gait detection in Parkinson's disease (PD) from short windows of triaxial acceleration, targeting resource-latency wearables and edge nodes.<n>We compare magnitude thresholding to three 1D CNNs for time-series analysis: a literature baseline (separable convolutions) and two ultra-light models - one purely separable and one with residual connections.
  arXiv  Detail & Related papers  (2025-11-29T08:52:41Z)
• Temporal Zoom Networks: Distance Regression and Continuous Depth for Efficient Action Localization [6.908972852063454]
  Temporal action localization requires both precise boundary detection and computational efficiency.<n>We address this through two complementary innovations: Boundary Distance Regression (BDR) and Adaptive Temporal Refinement (ATR)<n>On THUMOS14, our method achieves 56.5% mAP@0.7 with 151G FLOPs, using 36% fewer FLOPs than ActionFormer++ (55.7% mAP@0.7 at 235G)
  arXiv  Detail & Related papers  (2025-11-06T00:41:54Z)
• Environment-Aware Indoor LoRaWAN Path Loss: Parametric Regression Comparisons, Shadow Fading, and Calibrated Fade Margins [3.776919981139063]
  Indoor LoRaWAN propagation is shaped by structural and time-varying context factors.<n>We present an environment-aware, statistically disciplined path loss framework evaluated using leakage-safe cross-validation.
  arXiv  Detail & Related papers  (2025-10-05T20:14:48Z)
• Squeeze the Soaked Sponge: Efficient Off-policy Reinforcement Finetuning for Large Language Model [56.92219181993453]
  We propose Reincarnating Mix-policy Proximal Policy Gradient (ReMix) to enable on-policyRFT methods like PPO and GRPO to leverage off-policy data.<n>ReMix consists of three major components: (1) Mix-policy proximal policy gradient with an increased Update-To-Data (UTD) ratio for efficient training; (2) KL-Convex policy constraint to balance the trade-off between stability and flexibility; (3) Policy reincarnation to achieve a seamless transition from efficient early-stage learning to steady improvement.
  arXiv  Detail & Related papers  (2025-07-09T14:29:45Z)
• Tactile Grasp Refinement using Deep Reinforcement Learning and Analytic Grasp Stability Metrics [70.65363356763598]
  We show that analytic grasp stability metrics constitute powerful optimization objectives for reinforcement learning algorithms. We show that a combination of geometric and force-agnostic grasp stability metrics yields the highest average success rates of 95.4% for cuboids. In a second experiment, we show that grasp refinement algorithms trained with contact feedback perform up to 6.6% better than a baseline that receives no tactile information.
  arXiv  Detail & Related papers  (2021-09-23T09:20:19Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.