# paper-forge daily report — 2026-10-05

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

## Reproductions

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

## Top candidates

### [Does Learning Protein Folding Generalize to Broader Reasoning?](https://arxiv.org/abs/2609.38879)
`2609.38879` · score 54.7 · analysis: heuristic · feasibility 6/10

Large language models rely heavily on human text, which often conveys surface answers rather than the spatial and structural logic behind them. Protein folding is a natural testbed, because one solved structure yields thousands of exactly checkable spatial and topological statements.

Claimed: accuracy 45.09% (abstract (regex extraction, unverified))
Rubric evidence: - LLM-based ("Large language model"); + toy/synthetic experiments ("synthetic"); + small/synthetic data ("synthetic"); + public benchmarks ("benchmarks"); + code: https://github.com/GENTEL-lab/Fold2Reason; + 108 HF upvotes

### [MemFold: Learning Compact Soft Memory for Long-Context Personalization via On-Policy Optimization](https://arxiv.org/abs/2609.36435)
`2609.36435` · score 54.2 · analysis: heuristic · feasibility 5/10

An assistant that serves the same user over a long horizon has to answer from what that user has revealed: which preferences still hold, which were revised, and which constraints apply now. Retaining that information is not the same as acting on it, and the two are usually optimized as if they were.

Claimed: accuracy 32.0 (abstract (regex extraction, unverified))
Rubric evidence: + explicit algorithm/pseudocode ("Algorithm 1"); + code: https://github.com/Johnny221B/memfold; + 7 HF upvotes

### [SimuVerity: Benchmarking Agents for Engineering-Grade Simulink Model Generation](https://arxiv.org/abs/2610.02304)
`2610.02304` · score 53.7 · analysis: heuristic · feasibility 5/10

Existing Simulink benchmarks mainly evaluate whether generated models compile, execute, or resemble a reference model. These criteria do not establish whether a model satisfies its engineering requirements.

Rubric evidence: + toy/synthetic experiments ("simulation"); + public benchmarks ("benchmarks"); + code: https://github.com/SimuVerity/SimuVerity; + 21 HF upvotes

### [From Retrieval to Typed Decisions: Calibrated System One Models from Biomedical Sentence Encoders](https://arxiv.org/abs/2610.02486)
`2610.02486` · score 52.0 · analysis: heuristic · feasibility 5/10

Typed decision models answer schema-constrained questions about a text in one forward pass and return probabilities meant to be thresholded. We ask whether biomedical sentence encoders trained for retrieval are good starting points for such models.

Rubric evidence: + claims simplicity ("simple"); + code: https://github.com/pritamdeka/sbert2s1; + keywords: retrieval; + 2 HF upvotes

### [Science Utopia? Closed-Loop LLM Simulation of Academic Research Ecosystems](https://arxiv.org/abs/2610.01257)
`2610.01257` · score 51.6 · analysis: heuristic · feasibility 5/10

Scientific progress emerges from a longitudinal ecosystem in which researchers, institutions, funding agencies, collaboration networks, and the scientific literature co-evolve. As AI becomes increasingly involved throughout the scientific research cycle, understanding these interconnected and evolving processes becomes increasingly important.

Rubric evidence: - LLM-based ("LLM"); + toy/synthetic experiments ("Simulation"); + small/synthetic data ("simulated"); - system/framework (many moving parts) ("framework"); + code: https://github.com/Ahren09/ScienceUtopia; + 10 HF upvotes

### [Triadic Linear Attention: Three-Dimensional Recurrent States for Long-Context Sequence Modeling](https://arxiv.org/abs/2609.36529)
`2609.36529` · score 50.4 · analysis: heuristic · feasibility 5/10

Recurrent neural networks (RNNs) compress the historical context into a memory state of fixed size, thus allowing for constant-time inference. The memory state size is a crucial factor in their performance, as exemplified by the strong performance and resurgence of linear attention, which extends the vector-valued hidden states of ordinary RNNs to matrix-valued hidden states.

Rubric evidence: + formal statement ("bound"); + code: https://github.com/OliverSieberling/TriadicLinearAttention; + 5 HF upvotes

### [World Embedding Benchmark](https://arxiv.org/abs/2610.03632)
`2610.03632` · score 48.9 · analysis: heuristic · feasibility 4/10

Physical fidelity has received increasing attention in world models and video generation, yet how video representations encode physical information remains less understood. We introduce the World Embedding Benchmark, comprising 8,000 controlled simulation cases from 80 families spanning fluid mechanics, solid mechanics, dynamics, and optics & electromagnetism.

Rubric evidence: - pretraining ("pre-trained"); - diffusion/video ("video generation"); + toy/synthetic experiments ("simulation"); + CPU-friendly ("lightweight"); + public benchmarks ("Benchmark"); + code: https://github.com/World-Representation-Lab/World-Embedding-Benchmark; + keywords: retrieval; + 16 HF upvotes

### [GTR: Gated Token Recurrence for Efficient Dense Prediction](https://arxiv.org/abs/2609.26590)
`2609.26590` · score 46.3 · analysis: heuristic · feasibility 4/10

Self-attention-based vision backbones perform well on dense prediction, but the quadratic computational cost of global softmax attention limits their efficiency as image resolution increases. We introduce Gated Token Recurrence (GTR), a softmax-free recurrent vision backbone that combines gated linear attention, alternating spatial scan directions, and spatially enhanced SwiGLU blocks.

Rubric evidence: - pretraining ("pre-training"); + public benchmarks ("benchmark"); + claims simplicity ("simple"); + code: https://github.com/Intellindust-AI-Lab/GTR; + keywords: kernel; + 4 HF upvotes

### [Language Models that Play Chess and Explain Their Moves](https://arxiv.org/abs/2610.03695)
`2610.03695` · score 43.6 · analysis: heuristic · feasibility 4/10

Modern chess engines are silent experts: they play at a superhuman level, but do not offer explanations for their play. On the other hand, language models (LMs) can generate plausible-sounding explanations, but their weak playing strength limits the utility of their explanations.

Rubric evidence: - system/framework (many moving parts) ("framework"); + code: https://github.com/queen-project/queen; + 2 HF upvotes

### [HyperBrowseComp: A Multilingual and Multimodal Stress Test for Web-Browsing Agents](https://arxiv.org/abs/2610.03574)
`2610.03574` · score 41.6 · analysis: heuristic · feasibility 4/10

We introduce HyperBrowseComp, a multilingual and multimodal browsing benchmark comprising 423 manually authored and human-validated questions across 13 languages, written by native or highly proficient speakers. Questions are designed to be extremely challenging.

Rubric evidence: + public benchmarks ("benchmark"); no public code found; + keywords: retrieval; + 18 HF upvotes

## All screened papers

| score | paper | compute | data | clarity | code | interest | analysis | laptop-feasible |
|---:|---|---:|---:|---:|---:|---:|---|---|
| 54.7 | [Does Learning Protein Folding Generalize to Broader Reasoning?](https://arxiv.org/abs/2609.38879) `2609.38879` | 0.45 | 0.80 | 0.40 | 1.00 | 0.30 | heuristic | False |
| 54.2 | [MemFold: Learning Compact Soft Memory for Long-Context Personalization via On-Policy Optimization](https://arxiv.org/abs/2609.36435) `2609.36435` | 0.50 | 0.40 | 0.70 | 1.00 | 0.30 | heuristic | True |
| 53.7 | [SimuVerity: Benchmarking Agents for Engineering-Grade Simulink Model Generation](https://arxiv.org/abs/2610.02304) `2610.02304` | 0.65 | 0.45 | 0.40 | 1.00 | 0.30 | heuristic | True |
| 52.0 | [From Retrieval to Typed Decisions: Calibrated System One Models from Biomedical Sentence Encoders](https://arxiv.org/abs/2610.02486) `2610.02486` | 0.50 | 0.40 | 0.55 | 1.00 | 0.36 | heuristic | True |
| 51.6 | [Science Utopia? Closed-Loop LLM Simulation of Academic Research Ecosystems](https://arxiv.org/abs/2610.01257) `2610.01257` | 0.45 | 0.75 | 0.30 | 1.00 | 0.30 | heuristic | False |
| 50.4 | [Triadic Linear Attention: Three-Dimensional Recurrent States for Long-Context Sequence Modeling](https://arxiv.org/abs/2609.36529) `2609.36529` | 0.50 | 0.40 | 0.55 | 1.00 | 0.26 | heuristic | True |
| 48.9 | [World Embedding Benchmark](https://arxiv.org/abs/2610.03632) `2610.03632` | 0.40 | 0.45 | 0.40 | 1.00 | 0.50 | heuristic | False |
| 46.3 | [GTR: Gated Token Recurrence for Efficient Dense Prediction](https://arxiv.org/abs/2609.26590) `2609.26590` | 0.25 | 0.45 | 0.55 | 1.00 | 0.43 | heuristic | False |
| 43.6 | [Language Models that Play Chess and Explain Their Moves](https://arxiv.org/abs/2610.03695) `2610.03695` | 0.50 | 0.40 | 0.30 | 1.00 | 0.16 | heuristic | True |
| 41.6 | [HyperBrowseComp: A Multilingual and Multimodal Stress Test for Web-Browsing Agents](https://arxiv.org/abs/2610.03574) `2610.03574` | 0.50 | 0.45 | 0.40 | 0.00 | 0.50 | heuristic | True |
| 41.6 | [On-Policy Parameter Update Direction Underlies Generalization in LLM Post-Training](https://arxiv.org/abs/2609.36659) `2609.36659` | 0.30 | 0.40 | 0.40 | 1.00 | 0.30 | — | — |
| 41.6 | [Equal Ranking Quality, Different Decisions: Measuring and Reducing Order Dependence in LLM Scorers](https://arxiv.org/abs/2608.26762) `2608.26762` | 0.30 | 0.40 | 0.40 | 1.00 | 0.30 | — | — |
| 35.3 | [LexReward: A Taxonomy-Driven Reward Framework for Legal Language Models](https://arxiv.org/abs/2609.39071) `2609.39071` | 0.50 | 0.40 | 0.30 | 0.00 | 0.30 | — | — |
| 35.0 | [Looping Beyond Twice: A Scalable Recipe for Looped Mixture-of-Experts](https://arxiv.org/abs/2610.01153) `2610.01153` | 0.00 | 0.40 | 0.55 | 1.00 | 0.28 | — | — |
| 31.6 | [FrugalEvo: Towards Cost-Aware LLM-Guided Program Evolution](https://arxiv.org/abs/2610.03675) `2610.03675` | 0.30 | 0.40 | 0.00 | 1.00 | 0.20 | — | — |
| 31.6 | [Pivot-SD: Efficient Self-Distillation for Masked Diffusion Language Models](https://arxiv.org/abs/2610.03665) `2610.03665` | 0.35 | 0.45 | 0.30 | 0.00 | 0.30 | — | — |
| 31.1 | [Source Preference in the Wild: How LLM Agents Favor Items by Source, and How to Reduce It](https://arxiv.org/abs/2610.03195) `2610.03195` | 0.30 | 0.40 | 0.40 | 0.00 | 0.30 | — | — |
| 30.0 | [Multilingual GSM-Symbolic: What determines capability transfer across languages?](https://arxiv.org/abs/2610.03367) `2610.03367` | 0.00 | 0.40 | 0.30 | 1.00 | 0.30 | — | — |
| 30.0 | [Covert Assistance: Helpful LLM Agents Evade Oversight in Multi-Agent Systems](https://arxiv.org/abs/2609.39050) `2609.39050` | 0.30 | 0.40 | 0.40 | 0.00 | 0.23 | — | — |
| 29.9 | [Can Computation from Earlier Problems Help LLMs Solve New Ones?](https://arxiv.org/abs/2609.39394) `2609.39394` | 0.30 | 0.45 | 0.40 | 0.00 | 0.16 | — | — |
| 27.9 | [Video Generation Models: A Survey of Post-Training and Alignment](https://arxiv.org/abs/2610.00812) `2610.00812` | 0.10 | 0.45 | 0.00 | 1.00 | 0.30 | — | — |
| 26.7 | [Collective Bias Mitigation via Model Routing and Collaboration](https://arxiv.org/abs/2610.03240) `2610.03240` | 0.30 | 0.40 | 0.30 | 0.00 | 0.16 | — | — |
| 18.9 | [Rethinking Token Reweighting for SFT: Suppress, Reverse, and Extrapolate Learned Features](https://arxiv.org/abs/2609.33463) `2609.33463` | 0.25 | 0.45 | 0.00 | 0.00 | 0.10 | — | — |

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