LLM Foundation Models: February 2026 Week 7

1. Native Reasoning Models: Training Language Models to Reason on Unverifiable Data

arXiv:2602.11549

Why Novel: First comprehensive framework for training reasoning without external verifiers. Treats reasoning traces as latent variables with intrinsic rewards derived from predictive confidence.

Key Innovations:

• Formalizes reasoning as latent variable optimization: maximize $P(y^*|x,z)$ where $z$ is reasoning trace
• Introduces aggregation-based reward schemes: Geometric Mean, Weighted Sum with inverse-probability weighting
• NRT-WS$(-\log p)$ achieves 56.2% overall on Llama-3.1-8B vs 46.0% SFT (+10.2%)
• Prevents mode collapse: maintains high entropy, long traces (415 tokens vs 12 for baselines)
• Weighted schemes target model's uncertain tokens: largest gains on high-entropy predictions

Evidence:

• tab:1 — Intrinsic reward schemes with aggregation functions and token reward signals
• tab:3 — NRT-WS achieves 56.2% overall on Llama-3.1-8B across 9 benchmarks
• fig:2 — Training curves show NRT maintains entropy/length/quality while RLPR collapses
• fig:3 — Weighted schemes provide largest confidence gains on high-entropy tokens

Impact: Enables reasoning training on any SFT data without verifiers or expert traces. Opens path to scalable reasoning in unverifiable domains like creative writing, open-ended QA.

2. Amortized Reasoning Tree Search: Decoupling Proposal and Decision in Large Language Models

arXiv:2602.12846

Why Novel: First formal analysis of RLVR's 'Normalization Squeeze' that extinguishes rare but valid reasoning paths. Proposes flow-based verifier to preserve long-tail reasoning without modifying base model.

Key Innovations:

• Identifies 'Normalization Squeeze': RLVR acts as high-pass filter, suppressing rare valid traces exponentially
• Tracks relative log-likelihood: rare traces decay systematically despite being correct
• ARTS decouples Proposer (frozen base) from Verifier (flow-based) for inference-time search
• Flow Matching preserves diversity where discriminative verifiers saturate
• Matches GRPO finetuning (74.6% vs 74.7% BoN@16) with <0.3% trainable params

Evidence:

• fig:1 — Visualization of Normalization Squeeze: rare traces enter extinction zone under RLVR
• tab:1 — ARTS achieves 74.6% BoN@16 matching GRPO's 74.7% with frozen base model
• fig:3 — On extinction set (GRPO=0%), ARTS achieves 6.9% vs PRM's 1.7%
• tab:3 — ARTS 3.2x gain on rare solutions, 8x on Counting & Probability

Impact: Explains why RLVR struggles with hard problems: not capability but filtering. Provides practical alternative via inference-time search that preserves model diversity.

3. In-Context Function Learning in Large Language Models

arXiv:2602.11863

Why Novel: First principled framework for evaluating ICL on continuous functions via Gaussian Process regression. Shows LLMs achieve GP-like learning curves and priors can be steered via post-training.

Key Innovations:

• Casts ICL as GP regression with known priors: GP as lower bound, 1-NN as upper bound
• LLM learning curves approach GP baseline, well below 1-NN, with log-scaling in model size
• Inductive bias analysis via likelihood under different kernels reveals bias toward rough functions
• Bias shifts toward smoother functions in higher dimensions
• SFT and GRPO steer priors toward training data structure; GRPO offers better generalization

Evidence:

• fig:1 — Framework overview: ICL evaluation, post-training, and inductive bias analysis via GP likelihoods
• fig:2 — Learning curves: Qwen-3 approaches GP baseline, 14B/32B show log-scaling improvement
• fig:4 — Base models biased toward rough kernels (low $\nu$); predictions more likely under Matern-1/2
• fig:5 — Post-training shifts likelihood toward training data kernel

Impact: Provides quantitative framework for understanding and steering ICL priors. Enables principled design of post-training for data-efficient continuous-function tasks.

1. Prototype Transformer: Towards Language Model Architectures Interpretable by Design

arXiv:2602.11852

ProtoT replaces attention with prototype-based mixer achieving O(n) complexity. Prototypes learn nameable concepts, enable targeted edits, and support time-scale analysis for transparent reasoning.

2. The Pensieve Paradigm: Stateful Language Models Mastering Their Own Context

arXiv:2602.12108

StateLM learns to manage internal context via Pensieve-inspired external memory with read/write/delete ops. +52% on BrowseComp-Plus, +10-20% on chat memory via learned state management.

3. Think Fast and Slow: Step-Level Cognitive Depth Adaptation for LLM Agents

arXiv:2602.12662

CogRouter implements ACT-R-grounded step-level depth adaptation with four cognitive levels. CoSFT + CoPO training achieves SOTA on ALFWorld/ScienceWorld with substantially lower token usage.

4. ThinkRouter: Efficient Reasoning via Routing Thinking between Latent and Discrete Spaces

arXiv:2602.11683

Dynamic routing based on confidence $p_t^{\max}$ and threshold $\tau$. Training-free method yields +19.7pp Pass@1 on average by routing to discrete space when latent confidence is low.

5. Towards On-Policy SFT: Distribution Discriminant Theory and its Applications in LLM Training

arXiv:2602.12222

Distribution Discriminant Theory (DDT) quantifies data-model alignment. IDFT reweights SFT loss, Hinted Decoding realigns data, matching DPO/SimPO performance with SFT efficiency.