---
schema_version: "1.0.0"
edition_id: "mp-2026-10-12-morning-0094"
published_at: "2026-10-12T09:00:00.000-04:00"
modified_at: "2026-10-12T09:48:10.481-04:00"
canonical_url: "https://themachinepress.com/edition/2026-10-12"
story_count: 27
lead_story_id: "mp-2026-10-12-001"
---

# The Machine Press — Morning edition

Edition ID: `mp-2026-10-12-morning-0094`  
Published: 2026-10-12T09:00:00.000-04:00  
Canonical edition: https://themachinepress.com/edition/2026-10-12

A cryptographic construction shows why eliciting a model’s worst behavior cannot guarantee that every capability has been found.

## 1. An Auditor Can Open the Weights and Still Miss the Skill {#mp-2026-10-12-001}

- Story ID: `mp-2026-10-12-001`
- Type: `lead`
- Classification: `editorial`
- Content status: `updated`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-12-001/an-auditor-can-open-the-weights-and-still-miss-the-skill

**Dek:** A cryptographic construction shows why eliciting a model’s worst behavior cannot guarantee that every capability has been found.

Capability audits assume a hidden skill can eventually be brought into view. A new preprint challenges that assumption with deliberately constructed models that keep an encrypted neural network inside their weights and activate it only with a secret trigger. Under the authors’ cryptographic assumptions, the concealed computation remains difficult to discover even when an auditor can inspect all the weights. The team implemented proof-of-concept models in PyTorch and tested several elicitation methods, including fine-tuning and latent adversarial training, on a small chemical-reaction prediction task. The hidden behavior resisted those methods. This is a constructed adversarial case, not evidence that ordinary deployed models contain such mechanisms. Its significance is narrower and sharper: a clean elicitation test cannot, by itself, establish the absence of an intentionally concealed capability. The authors released their models to support defensive research.

### Why it matters {#why-it-matters-mp-2026-10-12-001}

A cryptographic construction shows why eliciting a model’s worst behavior cannot guarantee that every capability has been found.

### Limits and context {#limitations-mp-2026-10-12-001}

- A new preprint challenges that assumption with deliberately constructed models that keep an encrypted neural network inside their weights and activate it only with a secret trigger.
- This is a constructed adversarial case, not evidence that ordinary deployed models contain such mechanisms.
- Its significance is narrower and sharper: a clean elicitation test cannot, by itself, establish the absence of an intentionally concealed capability.

### Claims and sources {#claims-mp-2026-10-12-001}

- A cryptographic construction shows why eliciting a model’s worst behavior cannot guarantee that every capability has been found. [source-2026-10-12-001] — Qualification: A new preprint challenges that assumption with deliberately constructed models that keep an encrypted neural network inside their weights and activate it only with a secret trigger.

## 2. Protein Design Gets a Plain-Language Brief {#mp-2026-10-12-002}

- Story ID: `mp-2026-10-12-002`
- Type: `secondary`
- Classification: `editorial`
- Content status: `updated`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-12-002/protein-design-gets-a-plain-language-brief

**Dek:** Prompteina connects written design goals to all-atom structures, then tests a set of newly designed binders in the laboratory.

Protein designers usually need a different input representation for each constraint they want a model to honor. Prompteina makes natural language the interface to an all-atom structure generator. Its authors built a training collection of 1.8 million structures and more than 50 million captions describing global and residue-level properties. The model beat the evaluated baselines on 15 of 18 structural properties and at least 13 of 20 binder-design targets. The researchers also tested designs against GM2A, a target for which they report no previous de novo binders, and identified nine binders; the strongest had an apparent dissociation constant of 107 nanomolar. Those are bounded study results, not evidence of a treatment or a general ability to design any protein. The advance is an interface that can express a wider range of design objectives without requiring a separate handcrafted representation for every one.

### Why it matters {#why-it-matters-mp-2026-10-12-002}

Prompteina connects written design goals to all-atom structures, then tests a set of newly designed binders in the laboratory.

### Limits and context {#limitations-mp-2026-10-12-002}

- Those are bounded study results, not evidence of a treatment or a general ability to design any protein.

### Claims and sources {#claims-mp-2026-10-12-002}

- Prompteina connects written design goals to all-atom structures, then tests a set of newly designed binders in the laboratory. [source-2026-10-12-002] — Qualification: Those are bounded study results, not evidence of a treatment or a general ability to design any protein.

## 3. A Phone Scan Became the Robot’s Training Ground {#mp-2026-10-12-003}

- Story ID: `mp-2026-10-12-003`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `updated`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-12-003/a-phone-scan-became-the-robot-s-training-ground

**Dek:** Splat2Act builds scene-specific practice from captured rooms and the user’s own objects.

A generalist manipulation policy recorded zero pickup success in the authors’ captured deployment scenes before adaptation. Splat2Act uses a phone scan and object images to build a digital twin, discover reachable placement surfaces and generate successful simulated demonstrations. Across 101 campus environments, the pipeline produced more than 68,000 demonstrations. The strongest personalized policy reached 62.5% aggregate pickup success in its target scene. On a real Stretch 3 robot, it achieved 63.9% on a 12-task subset. The result supports scene-specific adaptation from simulated practice; it does not establish reliable manipulation in arbitrary homes.

### Why it matters {#why-it-matters-mp-2026-10-12-003}

Splat2Act builds scene-specific practice from captured rooms and the user’s own objects.

### Limits and context {#limitations-mp-2026-10-12-003}

- The result supports scene-specific adaptation from simulated practice; it does not establish reliable manipulation in arbitrary homes.

### Claims and sources {#claims-mp-2026-10-12-003}

- Splat2Act builds scene-specific practice from captured rooms and the user’s own objects. [source-2026-10-12-003] — Qualification: The result supports scene-specific adaptation from simulated practice; it does not establish reliable manipulation in arbitrary homes.

## 4. The Right Tool Also Had to Be Working {#mp-2026-10-12-004}

- Story ID: `mp-2026-10-12-004`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `updated`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-12-004/the-right-tool-also-had-to-be-working

**Dek:** RATED combines semantic matching with feedback about whether an agent’s tools actually execute.

Tool retrieval often asks which API matches a request while ignoring whether that API is available. RATED adds a small contextual-bandit reranker that learns from execution feedback without retraining the main retriever. Its EcoAgent benchmark includes 5,172 tools and more than 210,000 query-tool pairs. In the authors’ severe-outage evaluation, fault selection fell from 86.2% to 2.9%; they report less than one millisecond of added latency over encode-only inference. These are benchmark results under specified outage conditions, but they make operational availability part of the selection problem.

### Why it matters {#why-it-matters-mp-2026-10-12-004}

RATED combines semantic matching with feedback about whether an agent’s tools actually execute.

### Limits and context {#limitations-mp-2026-10-12-004}

- In the authors’ severe-outage evaluation, fault selection fell from 86.2% to 2.9%; they report less than one millisecond of added latency over encode-only inference.

### Claims and sources {#claims-mp-2026-10-12-004}

- RATED combines semantic matching with feedback about whether an agent’s tools actually execute. [source-2026-10-12-004] — Qualification: In the authors’ severe-outage evaluation, fault selection fell from 86.2% to 2.9%; they report less than one millisecond of added latency over encode-only inference.

## 5. The Generated Hinge Had to Move Without a Collision {#mp-2026-10-12-005}

- Story ID: `mp-2026-10-12-005`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `updated`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-12-005/the-generated-hinge-had-to-move-without-a-collision

**Dek:** Artisan tests how an articulated asset works, alongside how convincingly it looks.

A useful simulated cabinet needs more than a plausible surface. Artisan extends articulated 3D generation with multiple agents and physical-simulation feedback, checking collision-free motion, gravity stability, mass and materials. Its benchmark combines 100 reference assets, 50 stability prompts and 50 objects with reference masses. Against more than ten evaluated methods, the authors report competitive spatial fidelity alongside the lowest collision rates and highest stability rates. The paper’s target is a generated object that can function inside downstream tasks, rather than an attractive mesh whose mechanics were never tested.

### Why it matters {#why-it-matters-mp-2026-10-12-005}

Artisan tests how an articulated asset works, alongside how convincingly it looks.

### Limits and context {#limitations-mp-2026-10-12-005}

- No additional limitation was separately recorded.

### Claims and sources {#claims-mp-2026-10-12-005}

- Artisan tests how an articulated asset works, alongside how convincingly it looks. [source-2026-10-12-005]

## 6. Sophos Says Agent-Handled Cases Now Average 89 Seconds {#mp-2026-10-12-006}

- Story ID: `mp-2026-10-12-006`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `updated`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-12-006/sophos-says-agent-handled-cases-now-average-89-seconds

**Dek:** The October 9 case report puts speed gains inside the customer’s existing response permissions.

Sophos reports that cases handled with agents built through OpenAI Daybreak average about 89 seconds, compared with roughly 38 minutes in its previous process. The company says AI now resolves 52% of managed detection and response cases end to end. Investigation agents gather context and threat intelligence, then plan and summarize recommended responses. Customer settings still determine whether Sophos only notifies, collaborates before action, or responds directly under authorization. These figures come from a vendor case report, rather than an independent comparative trial. The deployment claim is faster case handling within defined permissions, with exceptions handed to human judgment.

### Why it matters {#why-it-matters-mp-2026-10-12-006}

The October 9 case report puts speed gains inside the customer’s existing response permissions.

### Limits and context {#limitations-mp-2026-10-12-006}

- The company says AI now resolves 52% of managed detection and response cases end to end.
- Customer settings still determine whether Sophos only notifies, collaborates before action, or responds directly under authorization.

### Claims and sources {#claims-mp-2026-10-12-006}

- The October 9 case report puts speed gains inside the customer’s existing response permissions. [source-2026-10-12-006] — Qualification: The company says AI now resolves 52% of managed detection and response cases end to end.

## 7. The Harm Appeared Between Otherwise Ordinary Tools {#mp-2026-10-12-007}

- Story ID: `mp-2026-10-12-007`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `updated`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-12-007/the-harm-appeared-between-otherwise-ordinary-tools

**Dek:** X-AgentRed searches service dependencies for risky combinations that a single-tool test can miss.

Personal agents can carry consequences from one service into another while fulfilling an ordinary request. X-AgentRed builds a graph of tool effects and execution preconditions, then uses feedback to explore and refine harmful cross-service paths. Across four frontier agentic models, its authors report 1.77 times the harmful-path rate of the STAC baseline and 2.99 times average tool coverage. This is an evaluation framework for discovering combinations of failures, not a count of incidents in production. The study argues that testing tools one at a time can leave their interactions underexplored.

### Why it matters {#why-it-matters-mp-2026-10-12-007}

X-AgentRed searches service dependencies for risky combinations that a single-tool test can miss.

### Limits and context {#limitations-mp-2026-10-12-007}

- This is an evaluation framework for discovering combinations of failures, not a count of incidents in production.

### Claims and sources {#claims-mp-2026-10-12-007}

- X-AgentRed searches service dependencies for risky combinations that a single-tool test can miss. [source-2026-10-12-007] — Qualification: This is an evaluation framework for discovering combinations of failures, not a count of incidents in production.

## 8. Four-Bit Inference Lost Its Calibration Dataset {#mp-2026-10-12-008}

- Story ID: `mp-2026-10-12-008`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `updated`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-12-008/four-bit-inference-lost-its-calibration-dataset

**Dek:** DualQuant derives activation scaling from the weight-reconstruction problem.

Activation quantization usually needs calibration examples even when compressing the weights does not. DualQuant uses scales obtained while minimizing weight-reconstruction error to rescale activations as well. It supports several low-bit formats, including MX and NVFP4, and the authors report matching or exceeding SmoothQuant accuracy across most tested models and formats. A packed four-bit weight-and-activation path for Llama-3.1-8B on an RTX 4090 achieved up to 2.61 times lower memory use and 2.46 times higher throughput than BF16. Those maxima belong to that evaluated implementation, rather than every model or accelerator.

### Why it matters {#why-it-matters-mp-2026-10-12-008}

DualQuant derives activation scaling from the weight-reconstruction problem.

### Limits and context {#limitations-mp-2026-10-12-008}

- Activation quantization usually needs calibration examples even when compressing the weights does not.

### Claims and sources {#claims-mp-2026-10-12-008}

- DualQuant derives activation scaling from the weight-reconstruction problem. [source-2026-10-12-008] — Qualification: Activation quantization usually needs calibration examples even when compressing the weights does not.

## 9. The Robot Learned When the Observation Was Taken {#mp-2026-10-12-009}

- Story ID: `mp-2026-10-12-009`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `updated`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-12-009/the-robot-learned-when-the-observation-was-taken

**Dek:** Chronos puts capture and execution times inside the policy instead of hiding them in sequence order.

Fixed-rate robot policies can become brittle when sensors slow down or inference takes longer. Chronos Policy explicitly conditions on observation-capture times and desired action-execution times. The authors compare timestamp embeddings with time-aware rotary positional encoding and find the latter more robust. Training randomizes sensing and execution schedules, allowing asynchronous measurements without forcing them onto one fixed grid. Simulation and real-world experiments show improvements on dynamic and multisensory tasks, while preserving nominal performance on the evaluated quasi-static tasks. Physical time becomes part of the model’s input, rather than an assumption made outside it.

### Why it matters {#why-it-matters-mp-2026-10-12-009}

Chronos puts capture and execution times inside the policy instead of hiding them in sequence order.

### Limits and context {#limitations-mp-2026-10-12-009}

- No additional limitation was separately recorded.

### Claims and sources {#claims-mp-2026-10-12-009}

- Chronos puts capture and execution times inside the policy instead of hiding them in sequence order. [source-2026-10-12-009]

## 10. A Useful Checkpoint Included More Than the Git Diff {#mp-2026-10-12-010}

- Story ID: `mp-2026-10-12-010`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `updated`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-12-010/a-useful-checkpoint-included-more-than-the-git-diff

**Dek:** Coding-agent recovery benefited when progress lived in databases and running services.

Restarting a failed coding attempt can throw away useful work outside the repository. This study implements Go-Explore-style search over full Daytona sandbox snapshots, restoring intermediate states and branching from them. Its authors find that higher and more efficient solve rates mainly depend on progress in databases, services and other runtime state. The finding is conditional: a full machine snapshot is most valuable when a long task spans state that a code diff cannot reconstruct. It is not a claim that snapshot search improves every repository-only coding problem.

### Why it matters {#why-it-matters-mp-2026-10-12-010}

Coding-agent recovery benefited when progress lived in databases and running services.

### Limits and context {#limitations-mp-2026-10-12-010}

- The finding is conditional: a full machine snapshot is most valuable when a long task spans state that a code diff cannot reconstruct.
- It is not a claim that snapshot search improves every repository-only coding problem.

### Claims and sources {#claims-mp-2026-10-12-010}

- Coding-agent recovery benefited when progress lived in databases and running services. [source-2026-10-12-010] — Qualification: The finding is conditional: a full machine snapshot is most valuable when a long task spans state that a code diff cannot reconstruct.

## 11. The Models Could Navigate the Encyclopedia. Humans Still Held Up. {#mp-2026-10-12-011}

- Story ID: `mp-2026-10-12-011`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `updated`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-12-011/the-models-could-navigate-the-encyclopedia-humans-still-held-up

**Dek:** A link-following game tests retrieval across five Wikipedia languages and five literature datasets.

The Wikipedia Game requires reaching a target article using only links from the current page. Researchers used local Wikipedia and academic-citation graphs to compare language models, embedding methods, people and TF-IDF baselines without relying on third-party navigation APIs. No model significantly outperformed humans, although the best evaluated models achieved comparable success. Rankings changed across languages and subject areas. The authors provide graph-building and navigation code, making the test a reproducible way to examine information-finding skills rather than another open-ended chat demonstration.

### Why it matters {#why-it-matters-mp-2026-10-12-011}

A link-following game tests retrieval across five Wikipedia languages and five literature datasets.

### Limits and context {#limitations-mp-2026-10-12-011}

- The Wikipedia Game requires reaching a target article using only links from the current page.

### Claims and sources {#claims-mp-2026-10-12-011}

- A link-following game tests retrieval across five Wikipedia languages and five literature datasets. [source-2026-10-12-011] — Qualification: The Wikipedia Game requires reaching a target article using only links from the current page.

## 12. The State-Space Server Remembered Inputs Instead of Rewriting State {#mp-2026-10-12-012}

- Story ID: `mp-2026-10-12-012`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `updated`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-12-012/the-state-space-server-remembered-inputs-instead-of-rewriting-state

**Dek:** ReplaySSM separates output computation from periodic checkpoint updates.

State-space models keep compact recurrent state, but moving that state on every decoding step can become expensive. ReplaySSM represents it with a checkpoint plus recent inputs, avoiding writes on most steps and making speculative rollback a buffer-pointer operation. Implemented in vLLM for Mamba-2 and Gated DeltaNet, it accelerated standard decoding by 1.20–1.47 times at batch size 256 across four hybrid models. On two models in the 120-billion-parameter class, speculative throughput was about 1.8 times the vLLM baseline. Separate checkpoint precisions also reduced the measured effects of quantization error.

### Why it matters {#why-it-matters-mp-2026-10-12-012}

ReplaySSM separates output computation from periodic checkpoint updates.

### Limits and context {#limitations-mp-2026-10-12-012}

- No additional limitation was separately recorded.

### Claims and sources {#claims-mp-2026-10-12-012}

- ReplaySSM separates output computation from periodic checkpoint updates. [source-2026-10-12-012]

## 13. Two Image Generators Shared a Place to Hand Off the Picture {#mp-2026-10-12-013}

- Story ID: `mp-2026-10-12-013`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `updated`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-12-013/two-image-generators-shared-a-place-to-hand-off-the-picture

**Dek:** InstantFusion links frozen generators through a learned common latent interface.

Independently trained image generators do not necessarily organize their internal states in identical ways. InstantFusion learns model-specific encoders and decoders that translate intermediate states through a shared space. Training aligns states from the same image at matched noise levels and synchronizes local denoising velocities. Once learned, the interface can switch between frozen generators without training a separate mapping for every model pair. The authors demonstrate uses in acceleration, preference composition and distillation. It is a proposed interoperability mechanism, with results limited to the generators and settings they evaluated.

### Why it matters {#why-it-matters-mp-2026-10-12-013}

InstantFusion links frozen generators through a learned common latent interface.

### Limits and context {#limitations-mp-2026-10-12-013}

- Independently trained image generators do not necessarily organize their internal states in identical ways.

### Claims and sources {#claims-mp-2026-10-12-013}

- InstantFusion links frozen generators through a learned common latent interface. [source-2026-10-12-013] — Qualification: Independently trained image generators do not necessarily organize their internal states in identical ways.

## 14. The Robot Stopped Without Waiting for Another Model Reply {#mp-2026-10-12-014}

- Story ID: `mp-2026-10-12-014`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `updated`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-12-014/the-robot-stopped-without-waiting-for-another-model-reply

**Dek:** BT-Agent pairs language-model decisions with locally checked behavior-tree skills.

BT-Agent runs a nine-billion-parameter model onboard a quadruped while predefined behavior-tree skills enforce execution conditions. In intervention tests, local checks stopped ongoing navigation before the language model made another decision; the agent could then retry or report the constraint. The study also introduces a 161-task benchmark, with 39 tasks evaluated in Habitat simulation. Its largest gains over the strongest Act/ReAct baseline were roughly ten percentage points at four billion parameters and five points at nine billion. Structured tool calls and useful failure feedback contributed most in the ablations.

### Why it matters {#why-it-matters-mp-2026-10-12-014}

BT-Agent pairs language-model decisions with locally checked behavior-tree skills.

### Limits and context {#limitations-mp-2026-10-12-014}

- No additional limitation was separately recorded.

### Claims and sources {#claims-mp-2026-10-12-014}

- BT-Agent pairs language-model decisions with locally checked behavior-tree skills. [source-2026-10-12-014]

## 15. The Robot’s World Model Fit a Smaller Compute Budget {#mp-2026-10-12-026}

- Story ID: `mp-2026-10-12-026`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `updated`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-12-026/the-robot-s-world-model-fit-a-smaller-compute-budget

**Dek:** SmolWAM compresses a pretrained policy’s inference path while retaining a future-scene head.

SmolWAM keeps a pretrained Cosmos Policy trunk but rebuilds the interface around single-frame visual latents, cached text, twelve transformer blocks and one rectified-flow step. The authors report 95.5% success across 400 LIBERO episodes, with 57-millisecond inference per sixteen-action chunk and 2.4 GiB of GPU memory. In their comparison it was eleven times faster than its teacher, with fewer operations and less memory, while still predicting scene motion. These are manipulation-benchmark results; the paper does not establish the same performance on every edge device or real-world task.

### Why it matters {#why-it-matters-mp-2026-10-12-026}

SmolWAM compresses a pretrained policy’s inference path while retaining a future-scene head.

### Limits and context {#limitations-mp-2026-10-12-026}

- These are manipulation-benchmark results; the paper does not establish the same performance on every edge device or real-world task.

### Claims and sources {#claims-mp-2026-10-12-026}

- SmolWAM compresses a pretrained policy’s inference path while retaining a future-scene head. [source-2026-10-12-015] — Qualification: These are manipulation-benchmark results; the paper does not establish the same performance on every edge device or real-world task.

## 16. The Model Learned to Withhold Some Answers. The Guarantee Did Not Travel. {#mp-2026-10-12-027}

- Story ID: `mp-2026-10-12-027`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `updated`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-12-027/the-model-learned-to-withhold-some-answers-the-guarantee-did-not-travel

**Dek:** A separate answer selector improved one factual dataset but still lost to silence on another.

Researchers asked whether reliable abstention should come from training the language model itself or from a separate component that checks its proposed answers. On Qwen3-14B, the learned selector improved threshold-averaged utility over the original model’s calibrated confidence filter. Under exact-match grading on TriviaQA, mean coverage rose from 20.2% to 30.9%, while the pooled error rate among scored answers fell from 23.0% to 14.5%. The tested reinforcement-learning changes did not produce consistent gains. On NQ-Open, the selector improved on the filter but still performed worse than always abstaining. A partial human-reviewed audit preserved the TriviaQA finding. The contrast matters for deploying a system that must decide whether it knows enough: selecting answers can help in a particular setting, but one dataset’s improvement does not establish reliable uncertainty handling elsewhere. The authors leave that broader problem unresolved.

### Why it matters {#why-it-matters-mp-2026-10-12-027}

A separate answer selector improved one factual dataset but still lost to silence on another.

### Limits and context {#limitations-mp-2026-10-12-027}

- The tested reinforcement-learning changes did not produce consistent gains.
- The contrast matters for deploying a system that must decide whether it knows enough: selecting answers can help in a particular setting, but one dataset’s improvement does not establish reliable uncertainty handling elsewhere.

### Claims and sources {#claims-mp-2026-10-12-027}

- A separate answer selector improved one factual dataset but still lost to silence on another. [source-2026-10-12-016] — Qualification: The tested reinforcement-learning changes did not produce consistent gains.

## 17. The Missing Object Was Harder to Read from Its Shadow {#mp-2026-10-12-015}

- Story ID: `mp-2026-10-12-015`
- Type: `ticker`
- Classification: `editorial`
- Content status: `updated`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-12-015/the-missing-object-was-harder-to-read-from-its-shadow

**Dek:** OccludeBench separates physical clues from broad scene context.

Across five vision-language models, the authors report 42.7% average accuracy when a hidden object must be recovered from a shadow or reflection, versus 62.7% from broader context. Human results were closer across the two groups. Reasoning and tool use did not consistently close the model gap.

### Why it matters {#why-it-matters-mp-2026-10-12-015}

OccludeBench separates physical clues from broad scene context.

### Limits and context {#limitations-mp-2026-10-12-015}

- Reasoning and tool use did not consistently close the model gap.

### Claims and sources {#claims-mp-2026-10-12-015}

- OccludeBench separates physical clues from broad scene context. [source-2026-10-12-017] — Qualification: Reasoning and tool use did not consistently close the model gap.

## 18. Imagined Practice Helped the Robot Keep Earlier Skills {#mp-2026-10-12-016}

- Story ID: `mp-2026-10-12-016`
- Type: `ticker`
- Classification: `editorial`
- Content status: `updated`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-12-016/imagined-practice-helped-the-robot-keep-earlier-skills

**Dek:** Dream to Remember rehearses prior tasks while a steering module learns new ones.

On eight sequential CALVIN tasks, the authors report average success rising from 63.7% to 74.8%; their LoRA comparison fell to 11.7%. The approach uses a frozen world model to rehearse earlier behavior, and also showed sequential improvements on three real Franka tasks.

### Why it matters {#why-it-matters-mp-2026-10-12-016}

Dream to Remember rehearses prior tasks while a steering module learns new ones.

### Limits and context {#limitations-mp-2026-10-12-016}

- No additional limitation was separately recorded.

### Claims and sources {#claims-mp-2026-10-12-016}

- Dream to Remember rehearses prior tasks while a steering module learns new ones. [source-2026-10-12-018]

## 19. The Token Merger Hit an Exact Budget in Linear Time {#mp-2026-10-12-017}

- Story ID: `mp-2026-10-12-017`
- Type: `ticker`
- Classification: `editorial`
- Content status: `updated`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-12-017/the-token-merger-hit-an-exact-budget-in-linear-time

**Dek:** LapMerge learns how to allocate an ordered sequence into a fixed number of outputs.

A learned mass distribution partitions tokens into equal-mass segments, with prefix and suffix scans doing the aggregation. The authors report competitive vision and vision-language results and large operator-speed gains on long sequences. The method works best when token order preserves spatial or temporal locality.

### Why it matters {#why-it-matters-mp-2026-10-12-017}

LapMerge learns how to allocate an ordered sequence into a fixed number of outputs.

### Limits and context {#limitations-mp-2026-10-12-017}

- No additional limitation was separately recorded.

### Claims and sources {#claims-mp-2026-10-12-017}

- LapMerge learns how to allocate an ordered sequence into a fixed number of outputs. [source-2026-10-12-019]

## 20. The Comics Archive Became a Study of Who Appears and Who Speaks {#mp-2026-10-12-018}

- Story ID: `mp-2026-10-12-018`
- Type: `ticker`
- Classification: `editorial`
- Content status: `updated`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-12-018/the-comics-archive-became-a-study-of-who-appears-and-who-speaks

**Dek:** PAGE combines machine annotation with a written codebook and human checks.

The framework parses scans into panels, text and bibliographic records, then grounds sampled model readings in human annotation. Its demonstrations span 3,929 American comic books and 87 manga volumes, examining political portrayal, visual form and dialogue. The paper proposes archive research tools rather than replacing human scholarship.

### Why it matters {#why-it-matters-mp-2026-10-12-018}

PAGE combines machine annotation with a written codebook and human checks.

### Limits and context {#limitations-mp-2026-10-12-018}

- No additional limitation was separately recorded.

### Claims and sources {#claims-mp-2026-10-12-018}

- PAGE combines machine annotation with a written codebook and human checks. [source-2026-10-12-020]

## 21. When Smoke Blinded LiDAR, Radar Took the Lead {#mp-2026-10-12-019}

- Story ID: `mp-2026-10-12-019`
- Type: `ticker`
- Classification: `editorial`
- Content status: `updated`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-12-019/when-smoke-blinded-lidar-radar-took-the-lead

**Dek:** A reliability estimate decides which localization system should have authority.

An LSTM autoencoder trained on nominal sensing detects departures from healthy LiDAR behavior. Tests on ground-robot data from underground tunnels showed adaptive switching to radar-inertial odometry and continued localization in dense smoke. The authors describe an open-source release as future work, not already available code.

### Why it matters {#why-it-matters-mp-2026-10-12-019}

A reliability estimate decides which localization system should have authority.

### Limits and context {#limitations-mp-2026-10-12-019}

- The authors describe an open-source release as future work, not already available code.

### Claims and sources {#claims-mp-2026-10-12-019}

- A reliability estimate decides which localization system should have authority. [source-2026-10-12-021] — Qualification: The authors describe an open-source release as future work, not already available code.

## 22. Space Drums 2.0 {#mp-2026-10-12-020}

- Story ID: `mp-2026-10-12-020`
- Type: `invention_desk`
- Classification: `editorial`
- Content status: `carried_over`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-12-020/space-drums-2-0

**Dek:** Moves motion tracking onto two ESP32-S3 drumsticks, adds haptic hits and a two-pedal hub, and sends velocity-sensitive drum events to open PC or Android playback software.

Moves motion tracking onto two ESP32-S3 drumsticks, adds haptic hits and a two-pedal hub, and sends velocity-sensitive drum events to open PC or Android playback software.

### Why it matters {#why-it-matters-mp-2026-10-12-020}

An independent builder is turning an improbable idea into a working project.

### Limits and context {#limitations-mp-2026-10-12-020}

- A Desk Pick is an editorial selection, not a product endorsement.

### Claims and sources {#claims-mp-2026-10-12-020}

- This Invention Desk entry makes no independently sourced news claim.

## 23. 3RIC {#mp-2026-10-12-021}

- Story ID: `mp-2026-10-12-021`
- Type: `invention_desk`
- Classification: `editorial`
- Content status: `carried_over`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-12-021/3ric

**Dek:** Pairs a chip-level 65C02 computer with custom hardware, video, sound, floppy, and microSD support with a cycle-honest emulator and browser assembler built around the same machine.

Pairs a chip-level 65C02 computer with custom hardware, video, sound, floppy, and microSD support with a cycle-honest emulator and browser assembler built around the same machine.

### Why it matters {#why-it-matters-mp-2026-10-12-021}

An independent builder is turning an improbable idea into a working project.

### Limits and context {#limitations-mp-2026-10-12-021}

- A Desk Pick is an editorial selection, not a product endorsement.

### Claims and sources {#claims-mp-2026-10-12-021}

- This Invention Desk entry makes no independently sourced news claim.

## 24. M3-CRETE {#mp-2026-10-12-022}

- Story ID: `mp-2026-10-12-022`
- Type: `invention_desk`
- Classification: `editorial`
- Content status: `carried_over`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-12-022/m3-crete

**Dek:** Publishes a pallet-scale cartesian motion reference, CAD, BOM, and low-voltage controls for supervised cementitious-extrusion research while explicitly excluding certified structural use.

Publishes a pallet-scale cartesian motion reference, CAD, BOM, and low-voltage controls for supervised cementitious-extrusion research while explicitly excluding certified structural use.

### Why it matters {#why-it-matters-mp-2026-10-12-022}

An independent builder is turning an improbable idea into a working project.

### Limits and context {#limitations-mp-2026-10-12-022}

- A Desk Pick is an editorial selection, not a product endorsement.

### Claims and sources {#claims-mp-2026-10-12-022}

- This Invention Desk entry makes no independently sourced news claim.

## 25. DeskNode {#mp-2026-10-12-023}

- Story ID: `mp-2026-10-12-023`
- Type: `invention_desk`
- Classification: `editorial`
- Content status: `carried_over`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-12-023/desknode

**Dek:** Turns a bare ESP32-S3 touchscreen into a USB-connected desk instrument for PC activity and optional room temperature, humidity, and light measurements.

Turns a bare ESP32-S3 touchscreen into a USB-connected desk instrument for PC activity and optional room temperature, humidity, and light measurements.

### Why it matters {#why-it-matters-mp-2026-10-12-023}

An independent builder is turning an improbable idea into a working project.

### Limits and context {#limitations-mp-2026-10-12-023}

- A Desk Pick is an editorial selection, not a product endorsement.

### Claims and sources {#claims-mp-2026-10-12-023}

- This Invention Desk entry makes no independently sourced news claim.

## 26. The First Paid Slot {#mp-2026-10-12-024}

- Story ID: `mp-2026-10-12-024`
- Type: `invention_desk`
- Classification: `house_example`
- Content status: `carried_over`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-12-024/the-first-paid-slot

**Dek:** A transparent preview of paid placement with one verified link and no claim of endorsement.

A transparent preview of paid placement with one verified link and no claim of endorsement.

House example - no advertiser paid. Payment will buy placement, never endorsement.

### Why it matters {#why-it-matters-mp-2026-10-12-024}

This placement explains how builders can appear in The Invention Desk without purchasing editorial endorsement.

### Limits and context {#limitations-mp-2026-10-12-024}

- House example - no advertiser paid. Payment will buy placement, never endorsement.

### Claims and sources {#claims-mp-2026-10-12-024}

- This Invention Desk entry makes no independently sourced news claim.

## 27. Put Your Project on the Desk {#mp-2026-10-12-025}

- Story ID: `mp-2026-10-12-025`
- Type: `invention_desk`
- Classification: `house_example`
- Content status: `carried_over`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-12-025/put-your-project-on-the-desk

**Dek:** One manually reviewed placement stays active for seven days and remains separate from Desk Picks.

One manually reviewed placement stays active for seven days and remains separate from Desk Picks.

Manual intake only. Payment buys placement, never endorsement, and every submission is reviewed.

### Why it matters {#why-it-matters-mp-2026-10-12-025}

This placement explains how builders can appear in The Invention Desk without purchasing editorial endorsement.

### Limits and context {#limitations-mp-2026-10-12-025}

- Manual intake only. Payment buys placement, never endorsement, and every submission is reviewed.

### Claims and sources {#claims-mp-2026-10-12-025}

- This Invention Desk entry makes no independently sourced news claim.

## Normalized sources

- **source-2026-10-12-001:** [arXiv preprint 2610.13928, original v1](https://arxiv.org/abs/2610.13928v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-12-002:** [arXiv preprint 2610.13917, original v1](https://arxiv.org/abs/2610.13917v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-12-003:** [arXiv preprint 2610.13945, original v1](https://arxiv.org/abs/2610.13945v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-12-004:** [arXiv preprint 2610.13889, original v1](https://arxiv.org/abs/2610.13889v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-12-005:** [arXiv preprint 2610.13947, original v1](https://arxiv.org/abs/2610.13947v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-12-006:** [OpenAI / Sophos, October 9 vendor case report](https://openai.com/index/sophos/) — OpenAI / Sophos; secondary_reporting
- **source-2026-10-12-007:** [arXiv preprint 2610.13886, original v1](https://arxiv.org/abs/2610.13886v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-12-008:** [arXiv preprint 2610.13767, original v1](https://arxiv.org/abs/2610.13767v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-12-009:** [arXiv preprint 2610.13929, original v1](https://arxiv.org/abs/2610.13929v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-12-010:** [arXiv preprint 2610.13841, original v1](https://arxiv.org/abs/2610.13841v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-12-011:** [arXiv preprint 2610.13948, original v1](https://arxiv.org/abs/2610.13948v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-12-012:** [arXiv preprint 2610.13778, original v1](https://arxiv.org/abs/2610.13778v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-12-013:** [arXiv preprint 2610.13944, original v1](https://arxiv.org/abs/2610.13944v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-12-014:** [arXiv preprint 2610.13862, original v1](https://arxiv.org/abs/2610.13862v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-12-015:** [arXiv preprint 2610.13740, original v1](https://arxiv.org/abs/2610.13740v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-12-016:** [arXiv preprint 2610.13821, original v1](https://arxiv.org/abs/2610.13821v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-12-017:** [arXiv preprint 2610.13790, original v1](https://arxiv.org/abs/2610.13790v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-12-018:** [arXiv preprint 2610.13816, original v1](https://arxiv.org/abs/2610.13816v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-12-019:** [arXiv preprint 2610.13858, original v1](https://arxiv.org/abs/2610.13858v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-12-020:** [arXiv preprint 2610.13840, original v1](https://arxiv.org/abs/2610.13840v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-12-021:** [arXiv preprint 2610.13788, original v1](https://arxiv.org/abs/2610.13788v1) — arXiv / submitting research authors; primary_research

## Corrections

- 2026-10-12T09:48:10.481-04:00: Publication note: The October 12, 2026 edition was prepared and published late on October 12 after the scheduled newsroom task could not start. The edition retains its October 12 date and 9:00 AM America/Detroit reveal time. Reports use sources available before that cutoff; actual preparation and image-generation times are preserved in the provenance records. (revision 2)