# paper-forge daily report — 2026-10-13

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

## Reproductions

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

## Top candidates

### [AgentSWE: Can Coding Agents Build the Agent You Actually Want?](https://arxiv.org/abs/2610.13361)
`2610.13361` · score 52.1 · analysis: heuristic · feasibility 5/10

As the production of software passes into the hands of software itself, coding agents are increasingly commissioned to build agents fitted to their users' own workflows. This delegation carries a quiet assumption, that the delivered agent will faithfully serve the requirement it was built for, rather than merely appear to.

Rubric evidence: + public benchmarks ("benchmark"); + formal statement ("theorem"); + code: https://github.com/VectorSpaceLab/AgentSWE; + 199 HF upvotes

### [Initialization Improves LLM-Driven Discovery](https://arxiv.org/abs/2610.00707)
`2610.00707` · score 49.9 · analysis: heuristic · feasibility 5/10

Large Language Models (LLMs) have been used for novel discovery of algorithms, theorems, drugs, and other tasks through the use of harnesses that prompt an LLM to iteratively optimize an objective. In this work, we study the relationship between the population of previous iterates and eventual discovery success.

Rubric evidence: - LLM-based ("LLM"); + concrete method ("we propose a simple"); + formal statement ("theorem"); + claims simplicity ("simple"); + code: https://github.com/msakarvadia/llm_optimizer; + 2 HF upvotes

### [ARCS: Towards Precise Text-to-SQL via Structured Disambiguation](https://arxiv.org/abs/2610.09396)
`2610.09396` · score 48.6 · analysis: heuristic · feasibility 4/10

As text-to-SQL systems move beyond demonstrations toward real-world deployment, ambiguity in user questions becomes a primary source of errors. Such ambiguities are often subtle, domain- or data-specific, and can silently cause system outputs to deviate from the user's true intent.

Claimed: accuracy 27.0% (abstract (regex extraction, unverified))
Rubric evidence: + public benchmarks ("benchmark"); + code: https://github.com/megagonlabs/tabulaflow; + 6 HF upvotes

### [Thinking Inertia: LLMs Keep Thinking When Told Not To](https://arxiv.org/abs/2610.11765)
`2610.11765` · score 44.7 · analysis: heuristic · feasibility 4/10

Large Language Models (LLMs) increasingly ship with explicit "thinking modes", yet their counterpart, "no-thinking", has received far less attention. We study LLMs' no-thinking behavior along two axes.

Rubric evidence: - LLM-based ("LLMs"); + claims simplicity ("simple"); + code: https://github.com/thinking-inertia/code; + 130 HF upvotes

### [Clinician use of language models diverges from how the models are evaluated](https://arxiv.org/abs/2610.11069)
`2610.11069` · score 42.1 · analysis: heuristic · feasibility 4/10

Large language model (LLM) assistants are being deployed to clinicians across health systems, and judgments about their readiness rest largely on benchmark scores, most of them derived from examination questions or curated cases. A benchmark predicts performance in deployment only to the extent that its items resemble real use, yet whether benchmarks reflect the work these systems receive has rarely been measured.

Rubric evidence: - LLM-based ("Large language model"); + public benchmarks ("benchmark"); - system/framework (many moving parts) ("framework"); + code: https://github.com/nyuolab/RCQ-Map; + keywords: retrieval; + 3 HF upvotes

### [Multi-Stream Attention Residuals](https://arxiv.org/abs/2610.13630)
`2610.13630` · score 36.3 · analysis: heuristic · feasibility 3/10

Attention Residuals (AttnRes) enables content-dependent retrieval over network depth, which can be viewed as single-stream softmax attention along the depth axis. Inspired by the success of multi-head attention along the sequence axis, we introduce Multi-Stream Attention Residuals (MSAR), an extension of AttnRes that generalizes this mechanism to multi-head depth attention, yielding substantial performance gains with minimal additional overhead.

Rubric evidence: - billion-scale model/data ("18B"); - pretraining ("pre-training"); + CPU-friendly ("lightweight"); + public benchmarks ("benchmarks"); + code: https://github.com/jianke0604/MSAR; + keywords: retrieval; + 33 HF upvotes

### [GroundingPI: A Grounding Foundation Model towards Physical Intelligence with Visual Primitives](https://arxiv.org/abs/2609.39601)
`2609.39601` · score 34.7 · analysis: heuristic · feasibility 3/10

Precise grounding matters. It specifies which object is the target and where that object is, even in clutter and for tiny objects, and it has to be fast enough for closed-loop control.

Rubric evidence: - pretraining ("pretraining"); - LLM-based ("Foundation Model"); + public benchmarks ("benchmarks"); + code: https://github.com/groundingpi/GroundingPI; + 191 HF upvotes

### [Few-Step Generation via Data-Space Iteration](https://arxiv.org/abs/2610.12102)
`2610.12102` · score 34.7 · analysis: heuristic · feasibility 4/10

Flow matching has emerged as a scalable paradigm for training high-quality generative models, but sampling from the learned probability flow requires many network evaluations. Distillation can reduce this cost to one or a few evaluations; however, one-step generation often sacrifices quality, making few-step generation the practical operating regime.

Rubric evidence: - web-scale data ("ImageNet"); + formal statement ("convergence"); + claims simplicity ("simple"); - system/framework (many moving parts) ("framework"); no public code found; + keywords: sampling; + 3 HF upvotes

### [GroundAnything: Reconciling Parallel Decoding with Precise Visual Grounding at Flash Speed](https://arxiv.org/abs/2609.39600)
`2609.39600` · score 33.2 · analysis: heuristic · feasibility 3/10

Autoregressive (AR) grounding models serialize spatial predictions, introducing sequential latency and imposing a causal order on output tokens. We view grounding as visual evidence extraction: objects, locations, and spatial relations are jointly constrained by the image and query, yet their dependencies do not imply an intrinsic left-to-right generation order.

Rubric evidence: - pretraining ("pretraining"); - LLM-based ("foundation model"); - diffusion/video ("diffusion"); + public benchmarks ("benchmarks"); + code: https://github.com/groundingpi/GroundAnything; + 20 HF upvotes

### [Scaling Agentic Data for Long-Horizon Terminal Intelligence](https://arxiv.org/abs/2610.13166)
`2610.13166` · score 20.5 · analysis: heuristic · feasibility 2/10

Terminal agents are moving beyond local operations toward completing real-world professional tasks. Short trajectories can demonstrate commands and tool calls, but they do not fully show how agents accumulate evidence through exploration, use it to inform judgments, continually adjust strategies based on feedback, and turn intermediate results into final deliverables.

Rubric evidence: - billion-scale model/data ("27B"); + public benchmarks ("benchmarks"); - system/framework (many moving parts) ("Agentic"); no public code found; + 129 HF upvotes

## All screened papers

| score | paper | compute | data | clarity | code | interest | analysis | laptop-feasible |
|---:|---|---:|---:|---:|---:|---:|---|---|
| 52.1 | [AgentSWE: Can Coding Agents Build the Agent You Actually Want?](https://arxiv.org/abs/2610.13361) `2610.13361` | 0.50 | 0.45 | 0.55 | 1.00 | 0.30 | heuristic | True |
| 49.9 | [Initialization Improves LLM-Driven Discovery](https://arxiv.org/abs/2610.00707) `2610.00707` | 0.30 | 0.40 | 0.90 | 1.00 | 0.16 | heuristic | False |
| 48.6 | [ARCS: Towards Precise Text-to-SQL via Structured Disambiguation](https://arxiv.org/abs/2610.09396) `2610.09396` | 0.50 | 0.45 | 0.40 | 1.00 | 0.28 | heuristic | True |
| 44.7 | [Thinking Inertia: LLMs Keep Thinking When Told Not To](https://arxiv.org/abs/2610.11765) `2610.11765` | 0.30 | 0.40 | 0.55 | 1.00 | 0.30 | heuristic | False |
| 42.1 | [Clinician use of language models diverges from how the models are evaluated](https://arxiv.org/abs/2610.11069) `2610.11069` | 0.30 | 0.45 | 0.30 | 1.00 | 0.40 | heuristic | False |
| 36.3 | [Multi-Stream Attention Residuals](https://arxiv.org/abs/2610.13630) `2610.13630` | 0.00 | 0.45 | 0.40 | 1.00 | 0.50 | heuristic | False |
| 34.7 | [GroundingPI: A Grounding Foundation Model towards Physical Intelligence with Visual Primitives](https://arxiv.org/abs/2609.39601) `2609.39601` | 0.05 | 0.45 | 0.40 | 1.00 | 0.30 | heuristic | False |
| 34.7 | [Few-Step Generation via Data-Space Iteration](https://arxiv.org/abs/2610.12102) `2610.12102` | 0.50 | 0.00 | 0.60 | 0.00 | 0.40 | heuristic | False |
| 33.2 | [GroundAnything: Reconciling Parallel Decoding with Precise Visual Grounding at Flash Speed](https://arxiv.org/abs/2609.39600) `2609.39600` | 0.00 | 0.45 | 0.40 | 1.00 | 0.30 | heuristic | False |
| 20.5 | [Scaling Agentic Data for Long-Horizon Terminal Intelligence](https://arxiv.org/abs/2610.13166) `2610.13166` | 0.00 | 0.45 | 0.30 | 0.00 | 0.30 | heuristic | False |

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