# paper-forge daily report — 2026-10-12

27 new papers screened; 10 analyzed in depth; 0 reproduction runs.

## Reproductions

- none today (reimplementation is opt-in: `forge daily --reimplement`)

## Top candidates

### [Continually Post-training Small Models via Synthetic Self-distillation](https://arxiv.org/abs/2610.13113)
`2610.13113` · score 60.0 · analysis: heuristic · feasibility 7/10

Continual post-training faces a dual challenge of acquisition and retention: models must learn new capabilities while preserving prior knowledge. While Self-Distillation Fine-Tuning (SDFT) mitigates this using self-supervision with in-context demonstrations, it relies on strong in-context learning and instruction-following abilities.

Rubric evidence: + toy/synthetic experiments ("Synthetic"); + small/synthetic data ("Synthetic"); + public benchmarks ("benchmarks"); + explicit algorithm/pseudocode ("Algorithm 1"); no public code found; + keywords: sampling; + 39 HF upvotes

### [EmbodiedCLBench: Evaluating Continual Learning for Self-Evolving Embodied Agents](https://arxiv.org/abs/2610.13021)
`2610.13021` · score 56.8 · analysis: heuristic · feasibility 6/10

Continual learning requires models to accumulate past experience and transfer it to new tasks. Despite substantial research on continual learning, existing work remains largely confined to non-embodied settings, without extending to embodied environments.

Rubric evidence: + toy/synthetic experiments ("simulation"); + public benchmarks ("benchmark"); + claims simplicity ("simple"); + code: https://github.com/PKU-VaLuE-Lab/EmbodiedCLBench; + 62 HF upvotes

### [MetaKernelBench: Measuring GPU Kernel Knowledge Transfer Beyond Code](https://arxiv.org/abs/2610.05014)
`2610.05014` · score 55.3 · analysis: heuristic · feasibility 5/10

Recent GPU kernel optimization agents retain what they learn in knowledge bases or as distilled skills. Kernel benchmarks score each attempt's implementation for correctness and speed but leave the reuse value of retained experience unmeasured.

Rubric evidence: + public benchmarks ("benchmarks"); + formal statement ("bound"); + code: https://github.com/YIGE24/MetaKernelBench; + keywords: kernel; + 12 HF upvotes

### [S-Space: Exploring Spatial Workspace in Multimodal Language Models](https://arxiv.org/abs/2610.13527)
`2610.13527` · score 55.3 · analysis: heuristic · feasibility 6/10

Multimodal language models (MLLMs) exhibit promising spatial reasoning capabilities, yet the internal mechanisms supporting them remain poorly understood. We uncover S-Space, a linear representation subspace in which object-token representations encode continuous coordinates along horizontal, vertical, and distance axes.

Rubric evidence: + concrete method ("We propose a simple"); + claims simplicity ("simple"); + code: https://github.com/mirros-lab/s-space; + 197 HF upvotes

### [PreSTO: Predictive Subtree Prefetching for Fast LLM Power Sampling](https://arxiv.org/abs/2610.13171)
`2610.13171` · score 54.2 · analysis: heuristic · feasibility 4/10

Power sampling improves large language model (LLM) reasoning at test time by sampling from a sharpened version of the base model's distribution. Its sampler, PowerMH, is slow because its Metropolis--Hastings (MH) updates are sequential: each accept/reject decision sets the state from which the next proposal is generated.

Rubric evidence: - LLM-based ("LLM"); + formal statement ("We prove"); + code: https://github.com/Nan-Jiang-Group/PreSTO; + keywords: sampling, kernel, monte carlo; + 9 HF upvotes

### [The Life Cycle of a Massive Activation: Stochastic Birth, Weight-Decay-Driven Growth, and Competitive Consolidation](https://arxiv.org/abs/2610.00423)
`2610.00423` · score 52.0 · analysis: heuristic · feasibility 5/10

Massive activations, residual-stream coordinates with magnitudes far larger than typical activations, are associated with attention sinks in transformers, but how their scale is regulated during training remains incompletely understood. Combining training-trajectory analyses and controlled interventions, we trace their emergence, growth, and consolidation.

Rubric evidence: + formal statement ("bound"); + code: https://github.com/apd-smcclendon/massive_activations_2; + keywords: optimizer; + 2 HF upvotes

### [AMLworld-Compact: Importance-Weighted Downsampling for Cost-Effective LLM Evaluation and Error Diagnosis](https://arxiv.org/abs/2610.13627)
`2610.13627` · score 51.1 · analysis: heuristic · feasibility 5/10

Evaluating large language models (LLMs) on AMLworld's smallest test splits is costly because each of over a million targets requires a separate graph prompt. Sampling reduces this cost, but small balanced subsets change unweighted precision and F1.

Claimed: recall 1.0 (abstract (regex extraction, unverified)); F1 0.0 (abstract (regex extraction, unverified))
Rubric evidence: - LLM-based ("LLM"); + explicit algorithm/pseudocode ("Algorithm 1"); + code: https://github.com/nat-nischw/AMLworld-Compact; + keywords: sampling; + 11 HF upvotes

### [RSI-Anything: Unlocking Recursive Self-Improvement with World Models](https://arxiv.org/abs/2610.13912)
`2610.13912` · score 50.0 · analysis: heuristic · feasibility 5/10

Recursive self-improvement (RSI) aims to transform interaction experience into persistent improvements, yet external feedback is costly and world-model predictions may become unreliable as policies evolve. We introduce RSI-Anything, a domain-general framework built on JEPA-Anything that uses predictive world models to guide action and learning-data selection.

Rubric evidence: + concrete method ("update rule"); - system/framework (many moving parts) ("framework"); + code: https://github.com/Gen-Verse/RSI-Anything; + 28 HF upvotes

### [Coding-Agent Benchmarks Should Match Their Users' Task Flows](https://arxiv.org/abs/2610.09633)
`2610.09633` · score 48.3 · analysis: heuristic · feasibility 4/10

The evaluation of coding agents generally strives to be as realistic as possible. In our study, we collect 4,782 agent sessions of real software engineers in JetBrains IDEs, which we call Production Sessions.

Rubric evidence: + public benchmarks ("Benchmarks"); + code: https://github.com/JetBrains-Research/SWE-TaskFlow-Generation-Public; + 5 HF upvotes

### [Rethinking Relative Policy Differences Transfer of On-Policy Distillation](https://arxiv.org/abs/2610.13205)
`2610.13205` · score 47.9 · analysis: heuristic · feasibility 4/10

On-policy distillation (OPD) trains a target policy on its own rollouts using token-level supervision from a stronger teacher, but typically relies on the teacher being substantially more capable than the target. Recent work relaxes this requirement by transferring the relative policy difference between an expert and a reference policy rather than the expert's absolute distribution.

Rubric evidence: + code: https://github.com/KnowledgeXLab/A2-OPD; + 10 HF upvotes

## All screened papers

| score | paper | compute | data | clarity | code | interest | analysis | laptop-feasible |
|---:|---|---:|---:|---:|---:|---:|---|---|
| 60.0 | [Continually Post-training Small Models via Synthetic Self-distillation](https://arxiv.org/abs/2610.13113) `2610.13113` | 0.65 | 0.80 | 0.70 | 0.00 | 0.50 | heuristic | True |
| 56.8 | [EmbodiedCLBench: Evaluating Continual Learning for Self-Evolving Embodied Agents](https://arxiv.org/abs/2610.13021) `2610.13021` | 0.65 | 0.45 | 0.55 | 1.00 | 0.30 | heuristic | True |
| 55.3 | [MetaKernelBench: Measuring GPU Kernel Knowledge Transfer Beyond Code](https://arxiv.org/abs/2610.05014) `2610.05014` | 0.50 | 0.45 | 0.55 | 1.00 | 0.50 | heuristic | True |
| 55.3 | [S-Space: Exploring Spatial Workspace in Multimodal Language Models](https://arxiv.org/abs/2610.13527) `2610.13527` | 0.50 | 0.40 | 0.75 | 1.00 | 0.30 | heuristic | True |
| 54.2 | [PreSTO: Predictive Subtree Prefetching for Fast LLM Power Sampling](https://arxiv.org/abs/2610.13171) `2610.13171` | 0.30 | 0.40 | 0.55 | 1.00 | 0.90 | heuristic | False |
| 52.0 | [The Life Cycle of a Massive Activation: Stochastic Birth, Weight-Decay-Driven Growth, and Competitive Consolidation](https://arxiv.org/abs/2610.00423) `2610.00423` | 0.50 | 0.40 | 0.55 | 1.00 | 0.36 | heuristic | True |
| 51.1 | [AMLworld-Compact: Importance-Weighted Downsampling for Cost-Effective LLM Evaluation and Error Diagnosis](https://arxiv.org/abs/2610.13627) `2610.13627` | 0.30 | 0.40 | 0.70 | 1.00 | 0.50 | heuristic | False |
| 50.0 | [RSI-Anything: Unlocking Recursive Self-Improvement with World Models](https://arxiv.org/abs/2610.13912) `2610.13912` | 0.50 | 0.40 | 0.50 | 1.00 | 0.30 | heuristic | True |
| 48.3 | [Coding-Agent Benchmarks Should Match Their Users' Task Flows](https://arxiv.org/abs/2610.09633) `2610.09633` | 0.50 | 0.45 | 0.40 | 1.00 | 0.26 | heuristic | True |
| 47.9 | [Rethinking Relative Policy Differences Transfer of On-Policy Distillation](https://arxiv.org/abs/2610.13205) `2610.13205` | 0.50 | 0.40 | 0.40 | 1.00 | 0.30 | heuristic | True |
| 43.7 | [Inducing Process Supervision from Outcome-Only Reinforcement Learning](https://arxiv.org/abs/2609.36641) `2609.36641` | 0.30 | 0.45 | 0.30 | 1.00 | 0.50 | — | — |
| 42.6 | [Contextual Correction of Parallel Draftsfor Speculative Decoding](https://arxiv.org/abs/2610.12570) `2610.12570` | 0.30 | 0.45 | 0.40 | 1.00 | 0.30 | — | — |
| 42.6 | [Encore: Few-Shot Agentic Discovery of Manipulation Strategies](https://arxiv.org/abs/2609.37359) `2609.37359` | 0.50 | 0.40 | 0.30 | 1.00 | 0.10 | — | — |
| 40.5 | [SWE-Duel: Towards a Self-Scaling Arena for Coding Agents](https://arxiv.org/abs/2610.12569) `2610.12569` | 0.50 | 0.45 | 0.00 | 1.00 | 0.30 | — | — |
| 40.4 | [QuantMLA: Function-Aligned Dual-Path Quantization for Low-Bit MLA KV Caching](https://arxiv.org/abs/2609.36760) `2609.36760` | 0.50 | 0.45 | 0.45 | 0.00 | 0.36 | — | — |
| 36.8 | [Self-Preservation in Language Models](https://arxiv.org/abs/2610.12602) `2610.12602` | 0.25 | 0.40 | 0.40 | 1.00 | 0.10 | — | — |
| 36.3 | [Recurrent Harness for Agentic Continual Learning](https://arxiv.org/abs/2610.12668) `2610.12668` | 0.50 | 0.45 | 0.30 | 0.00 | 0.30 | — | — |
| 36.3 | [VisionWeave: Weaving Elastic Visual Representations as a Native Capability of MLLMs](https://arxiv.org/abs/2610.07987) `2610.07987` | 0.00 | 0.45 | 0.40 | 1.00 | 0.50 | — | — |
| 35.8 | [Which Language Should a Skeleton Speak? Language Choices in Multilingual Reasoning](https://arxiv.org/abs/2610.09607) `2610.09607` | 0.30 | 0.45 | 0.30 | 1.00 | 0.00 | — | — |
| 34.2 | [Stateless Language Agents: Scaling Long-Horizon Automated Research](https://arxiv.org/abs/2610.07625) `2610.07625` | 0.00 | 0.45 | 0.30 | 1.00 | 0.50 | — | — |
| 32.1 | [Surfing the Activation Wave: Analyzing the Information and Geometry of LLM Hidden States for Speculative Decoding](https://arxiv.org/abs/2610.12939) `2610.12939` | 0.30 | 0.40 | 0.30 | 0.00 | 0.50 | — | — |
| 32.1 | [DiffusionART: Efficient, Stable and Scalable Reinforcement of Diffusion Models](https://arxiv.org/abs/2610.12572) `2610.12572` | 0.00 | 0.40 | 0.40 | 1.00 | 0.30 | — | — |
| 32.1 | [When Do Biological Reasoning Models Use Their Biological Inputs?](https://arxiv.org/abs/2610.00898) `2610.00898` | 0.30 | 0.45 | 0.40 | 0.00 | 0.30 | — | — |
| 31.0 | [Distributionally Robust Mixture-of-Experts Training](https://arxiv.org/abs/2610.07207) `2610.07207` | 0.00 | 0.40 | 0.40 | 1.00 | 0.23 | — | — |
| 27.6 | [When Do We Need On-Policy Distillation? Distilling on Offline Student Rollouts Is Often Better](https://arxiv.org/abs/2610.11291) `2610.11291` | 0.00 | 0.40 | 0.55 | 0.00 | 0.48 | — | — |
| 21.6 | [One Model to Loop Them All: Elastic Latent Recurrence for Language Models](https://arxiv.org/abs/2610.12582) `2610.12582` | 0.00 | 0.40 | 0.40 | 0.00 | 0.30 | — | — |
| 21.1 | [SAGE: Salient Factor Discovery and Generation with Visual Foundation Representations](https://arxiv.org/abs/2609.39635) `2609.39635` | 0.35 | 0.00 | 0.40 | 0.00 | 0.10 | — | — |

Scores: weighted mean of rubric criteria (0–1 each) × 100; see README for the rubric.