---
schema_version: "1.0.0"
edition_id: "mp-2026-10-16-morning-0098"
published_at: "2026-10-16T09:00:00.000-04:00"
modified_at: "2026-10-16T09:00:00.000-04:00"
canonical_url: "https://themachinepress.com/edition/2026-10-16"
story_count: 27
lead_story_id: "mp-2026-10-16-001"
---

# The Machine Press — Morning edition

Edition ID: `mp-2026-10-16-morning-0098`  
Published: 2026-10-16T09:00:00.000-04:00  
Canonical edition: https://themachinepress.com/edition/2026-10-16

A scan of 21,097 MCP servers finds a permission gap: approving one declared tool can leave its process able to reach files outside that contract.

## 1. The Tool Contract Ends Before the Filesystem Does {#mp-2026-10-16-001}

- Story ID: `mp-2026-10-16-001`
- Type: `lead`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-16-001/the-tool-contract-ends-before-the-filesystem-does

**Dek:** A scan of 21,097 MCP servers finds a permission gap: approving one declared tool can leave its process able to reach files outside that contract.

Model Context Protocol clients commonly ask a person to approve a named tool with a description and argument schema. The underlying local server process, however, can retain much broader host permissions. The MCP-FSD study calls the resulting gap filesystem capability laundering: attacker-steered arguments lead a tool beyond its declared operation and deployment scope. Across 21,097 servers from nine registries, the authors triaged 4,806 and found 787 with filesystem capability despite non-filesystem classifications; paired dynamic tests confirmed 150 vulnerable instances under the tested sandbox policies. Coordinated disclosure produced 13 advisories, including two CVEs in Anthropic's Git MCP Server. These are the paper's measurements, not evidence that every MCP server or client is exploitable. The operational lesson is precise: a tool-call approval is not a least-privilege boundary unless the process beneath it is actually constrained.

### Why it matters {#why-it-matters-mp-2026-10-16-001}

A scan of 21,097 MCP servers finds a permission gap: approving one declared tool can leave its process able to reach files outside that contract.

### Limits and context {#limitations-mp-2026-10-16-001}

- These are the paper's measurements, not evidence that every MCP server or client is exploitable.
- The operational lesson is precise: a tool-call approval is not a least-privilege boundary unless the process beneath it is actually constrained.

### Claims and sources {#claims-mp-2026-10-16-001}

- A scan of 21,097 MCP servers finds a permission gap: approving one declared tool can leave its process able to reach files outside that contract. [source-2026-10-16-001] — Qualification: These are the paper's measurements, not evidence that every MCP server or client is exploitable.

## 2. Finishing the Chore Is Not the Same as Doing It Your Way {#mp-2026-10-16-002}

- Story ID: `mp-2026-10-16-002`
- Type: `secondary`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-16-002/finishing-the-chore-is-not-the-same-as-doing-it-your-way

**Dek:** LIBERO-PeRM separates task success from preference satisfaction across 312 simulated manipulation tasks—and current policies struggle before adaptation.

A household robot can put an object in the right place and still ignore how its user wanted the task done. LIBERO-PeRM turns that distinction into a controlled benchmark spanning motion, skill and user-profile preferences across five suites and 312 tasks. The authors test zero-shot following, learning from demonstrations, data efficiency, generalization, composition and conflict handling. Current vision-language-action policies struggled with zero-shot personalization, while supervised fine-tuning made preferences learnable. Mixed training preserved more of the no-preference behavior than sequential adaptation and benefited from contextual prompts, but it could weaken motion-preference learning; generalization and conflicting preferences remained difficult. The benchmark is simulated and limited to language-expressible preferences that can be evaluated automatically, so it does not establish safe household deployment. Its useful contribution is to stop counting mere completion as proof that a robot served the person using it.

### Why it matters {#why-it-matters-mp-2026-10-16-002}

LIBERO-PeRM separates task success from preference satisfaction across 312 simulated manipulation tasks—and current policies struggle before adaptation.

### Limits and context {#limitations-mp-2026-10-16-002}

- The benchmark is simulated and limited to language-expressible preferences that can be evaluated automatically, so it does not establish safe household deployment.

### Claims and sources {#claims-mp-2026-10-16-002}

- LIBERO-PeRM separates task success from preference satisfaction across 312 simulated manipulation tasks—and current policies struggle before adaptation. [source-2026-10-16-002] — Qualification: The benchmark is simulated and limited to language-expressible preferences that can be evaluated automatically, so it does not establish safe household deployment.

## 3. A Video Generator Becomes a One-Pass 4D Observer {#mp-2026-10-16-003}

- Story ID: `mp-2026-10-16-003`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-16-003/a-video-generator-becomes-a-one-pass-4d-observer

**Dek:** Gen4P reuses clean video latents and an on-demand decoder to estimate geometry, motion, depth and camera pose without iterative denoising.

Video generators learn more than how frames look. Gen4P feeds clean video latents through a pretrained generator's encoder once, then uses lightweight cross-attention queries to request 4D attributes at chosen places and times. A frozen video-depth adapter supplies extra geometric prior without new trainable parameters. The authors report state-of-the-art or competitive results for 3D point tracking, video depth and camera-pose estimation, plus qualitative motion tracking on in-the-wild clips. Those are research evaluations, not a general claim that generated-video models understand every physical scene.

### Why it matters {#why-it-matters-mp-2026-10-16-003}

Gen4P reuses clean video latents and an on-demand decoder to estimate geometry, motion, depth and camera pose without iterative denoising.

### Limits and context {#limitations-mp-2026-10-16-003}

- Those are research evaluations, not a general claim that generated-video models understand every physical scene.

### Claims and sources {#claims-mp-2026-10-16-003}

- Gen4P reuses clean video latents and an on-demand decoder to estimate geometry, motion, depth and camera pose without iterative denoising. [source-2026-10-16-003] — Qualification: Those are research evaluations, not a general claim that generated-video models understand every physical scene.

## 4. A Thousand Hours of Touch Improves What the Robot Sees {#mp-2026-10-16-004}

- Story ID: `mp-2026-10-16-004`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-16-004/a-thousand-hours-of-touch-improves-what-the-robot-sees

**Dek:** Spark-0 uses glove-derived contact maps and tactile histories during training, then supports both vision-only and touch-conditioned deployment.

Contact-rich manipulation hides force, compliance and incipient slip from ordinary cameras. Spark-0 pretrains visual tokens with contact maps from tactile gloves and a frozen 3D geometry teacher, while a separate path encodes force histories and taxel states. The authors built a 1,000-hour HumanTouch dataset and evaluated the model in simulation and on a real dexterous hand. Against vision-only pretraining on the same data and compute, keeping tactile supervision raised reported Ego Humanoid success from 52.5 to 71.7 percent; using touch at deployment raised SparkArena-Real success from 44 to 52 percent. These are benchmark results on the stated embodiments, not proof of general dexterity.

### Why it matters {#why-it-matters-mp-2026-10-16-004}

Spark-0 uses glove-derived contact maps and tactile histories during training, then supports both vision-only and touch-conditioned deployment.

### Limits and context {#limitations-mp-2026-10-16-004}

- Against vision-only pretraining on the same data and compute, keeping tactile supervision raised reported Ego Humanoid success from 52.5 to 71.7 percent; using touch at deployment raised SparkArena-Real success from 44 to 52 percent.
- These are benchmark results on the stated embodiments, not proof of general dexterity.

### Claims and sources {#claims-mp-2026-10-16-004}

- Spark-0 uses glove-derived contact maps and tactile histories during training, then supports both vision-only and touch-conditioned deployment. [source-2026-10-16-004] — Qualification: Against vision-only pretraining on the same data and compute, keeping tactile supervision raised reported Ego Humanoid success from 52.5 to 71.7 percent; using touch at deployment raised SparkArena-Real success from 44 to 52 percent.

## 5. The Best Agent Delivered 14 of 95 Customer Jobs {#mp-2026-10-16-005}

- Story ID: `mp-2026-10-16-005`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-16-005/the-best-agent-delivered-14-of-95-customer-jobs

**Dek:** FDE-Gym replaces tidy tickets with ambiguous requirements, scripted stakeholders and production-like systems—and scores the customer's outcome.

Agent benchmarks usually begin after somebody has already turned a messy need into a clean task. FDE-Gym restores the missing work: 95 delivery projects from 20 customer scenarios across 18 sectors require models to question stakeholders, modify live-like systems and uncover unstated constraints. The best of 13 evaluated configurations averaged 0.56 on the benchmark's normalized business-uplift score and cleared the authors' success threshold on only 14 tasks. Even when it found every key fact on 23 tasks, it reached at least 0.8 uplift on just 10. The simulated customers and grading system do not reproduce every deployment, but the result isolates a real distinction between solving a ticket and delivering a useful system.

### Why it matters {#why-it-matters-mp-2026-10-16-005}

FDE-Gym replaces tidy tickets with ambiguous requirements, scripted stakeholders and production-like systems—and scores the customer's outcome.

### Limits and context {#limitations-mp-2026-10-16-005}

- The best of 13 evaluated configurations averaged 0.56 on the benchmark's normalized business-uplift score and cleared the authors' success threshold on only 14 tasks.
- The simulated customers and grading system do not reproduce every deployment, but the result isolates a real distinction between solving a ticket and delivering a useful system.

### Claims and sources {#claims-mp-2026-10-16-005}

- FDE-Gym replaces tidy tickets with ambiguous requirements, scripted stakeholders and production-like systems—and scores the customer's outcome. [source-2026-10-16-005] — Qualification: The best of 13 evaluated configurations averaged 0.56 on the benchmark's normalized business-uplift score and cleared the authors' success threshold on only 14 tasks.

## 6. A Machine-Learning Agent Learns Between Experiments {#mp-2026-10-16-006}

- Story ID: `mp-2026-10-16-006`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-16-006/a-machine-learning-agent-learns-between-experiments

**Dek:** MLExpert assigns shared rewards to the decisions between costly experiment results instead of waiting for an entire engineering run to end.

Long machine-learning engineering runs make supervision sparse: one result reflects several earlier decisions, and each experiment can be costly. MLExpert is a 35-billion-parameter mixture-of-experts model trained with hierarchical supervised tuning and Windowed Interaction RL, which treats the decisions between successive experiment results as one rewarded window. On MLE-Bench Lite, the authors report that the model's average medal rate rose from 37.9 to 72.7 percent, with further gains on scientific and algorithm-optimization tasks. The comparisons are benchmark-specific and do not demonstrate recursive self-improvement. They do show a practical way to assign credit before a long experimental trajectory is finished.

### Why it matters {#why-it-matters-mp-2026-10-16-006}

MLExpert assigns shared rewards to the decisions between costly experiment results instead of waiting for an entire engineering run to end.

### Limits and context {#limitations-mp-2026-10-16-006}

- The comparisons are benchmark-specific and do not demonstrate recursive self-improvement.

### Claims and sources {#claims-mp-2026-10-16-006}

- MLExpert assigns shared rewards to the decisions between costly experiment results instead of waiting for an entire engineering run to end. [source-2026-10-16-006] — Qualification: The comparisons are benchmark-specific and do not demonstrate recursive self-improvement.

## 7. Editable 3D Worlds Move Into the Coding Harness {#mp-2026-10-16-007}

- Story ID: `mp-2026-10-16-007`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-16-007/editable-3d-worlds-move-into-the-coding-harness

**Dek:** 3DCodeVerse pairs two million executable 3D programs with automated execution, spatial checks and visual feedback.

A mesh can look plausible yet remain hard to edit, verify or reproduce. 3DCodeVerse represents 3D content as executable code packaged with the source files and assets needed to run it. Its two-million-program corpus feeds a multi-agent harness that executes projects, measures spatial properties and uses visual feedback to revise them. The authors report that fine-tuning improves open models and that their harness, when driven by Claude Opus 5.5, won 90 percent of pairwise human comparisons against Claude Design's 3D-object mode. That comparison covers the described evaluation, not every 3D workflow, but it makes editability and execution part of the artifact rather than an afterthought.

### Why it matters {#why-it-matters-mp-2026-10-16-007}

3DCodeVerse pairs two million executable 3D programs with automated execution, spatial checks and visual feedback.

### Limits and context {#limitations-mp-2026-10-16-007}

- A mesh can look plausible yet remain hard to edit, verify or reproduce.
- That comparison covers the described evaluation, not every 3D workflow, but it makes editability and execution part of the artifact rather than an afterthought.

### Claims and sources {#claims-mp-2026-10-16-007}

- 3DCodeVerse pairs two million executable 3D programs with automated execution, spatial checks and visual feedback. [source-2026-10-16-007] — Qualification: A mesh can look plausible yet remain hard to edit, verify or reproduce.

## 8. Fault-Tolerant Quantum Programs Get Assertions {#mp-2026-10-16-008}

- Story ID: `mp-2026-10-16-008`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-16-008/fault-tolerant-quantum-programs-get-assertions

**Dek:** SynDebug places lightweight checks inside encoded execution flows with 1.8 percent average added circuit depth in the authors' tests.

Quantum error correction adds layers of systems behavior that ordinary logical-circuit debuggers do not see. SynDebug introduces assertion-style primitives that operate inside fault-tolerant encoded infrastructures and test state preparation, entanglement and Shor-style programs. Across the evaluated code families and distances, the authors report 1.8 percent average additional circuit depth, zero overhead in most cases and no measurable increase in logical error rate. These are simulation and benchmark results, not evidence from a production-scale fault-tolerant computer. The contribution is architectural: debug checks can live inside the protected execution path instead of outside it.

### Why it matters {#why-it-matters-mp-2026-10-16-008}

SynDebug places lightweight checks inside encoded execution flows with 1.8 percent average added circuit depth in the authors' tests.

### Limits and context {#limitations-mp-2026-10-16-008}

- Quantum error correction adds layers of systems behavior that ordinary logical-circuit debuggers do not see.
- These are simulation and benchmark results, not evidence from a production-scale fault-tolerant computer.

### Claims and sources {#claims-mp-2026-10-16-008}

- SynDebug places lightweight checks inside encoded execution flows with 1.8 percent average added circuit depth in the authors' tests. [source-2026-10-16-008] — Qualification: Quantum error correction adds layers of systems behavior that ordinary logical-circuit debuggers do not see.

## 9. The Agent Organizes the Library Before the Question Arrives {#mp-2026-10-16-009}

- Story ID: `mp-2026-10-16-009`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-16-009/the-agent-organizes-the-library-before-the-question-arrives

**Dek:** Librarian Agents train a builder to turn raw documents into reusable file hierarchies optimized for later answering.

Retrieval normally waits for a question, leaving cross-document organization to inference time. Librarian Agents instead have a builder model create a hierarchical workspace of synthetic files and reference links before future queries are known. Reinforcement learning rewards structures that improve a separate answering agent. The authors report a 33-point gain over their strongest retrieval and precomputation baselines on BrowseComp-Plus and a 31-point average zero-shot gain on BRIGHT, with upfront construction costs amortized over repeated queries. The benchmarks do not guarantee faithful curation in every domain, but they test an important shift from finding documents to designing a reusable information environment.

### Why it matters {#why-it-matters-mp-2026-10-16-009}

Librarian Agents train a builder to turn raw documents into reusable file hierarchies optimized for later answering.

### Limits and context {#limitations-mp-2026-10-16-009}

- The benchmarks do not guarantee faithful curation in every domain, but they test an important shift from finding documents to designing a reusable information environment.

### Claims and sources {#claims-mp-2026-10-16-009}

- Librarian Agents train a builder to turn raw documents into reusable file hierarchies optimized for later answering. [source-2026-10-16-009] — Qualification: The benchmarks do not guarantee faithful curation in every domain, but they test an important shift from finding documents to designing a reusable information environment.

## 10. A Smaller Agent Decides When to Call the Expensive One {#mp-2026-10-16-010}

- Story ID: `mp-2026-10-16-010`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-16-010/a-smaller-agent-decides-when-to-call-the-expensive-one

**Dek:** LEASH uses latent uncertainty for routing, step-level escalation and a shadow observer after a strong black-box agent takes over.

Using a frontier model at every computer-use step is costly, but leaving the whole task to a smaller model often fails. LEASH compares a weak model's hidden state with its past successful and failed trajectories, escalating when the state resembles failure. After handoff, the weak model continues as a shadow observer because the stronger API exposes no internals. On OSWorld across four backbones, the authors report gains of 6.55 to 11.94 points over random escalation at matched rates. One configuration reached 73.66 percent task success at $0.4117 per task, roughly matching the paper's always-on Claude Opus 5 result while reducing inference cost 19.2 percent. These figures are benchmark- and pricing-specific, not a deployment guarantee.

### Why it matters {#why-it-matters-mp-2026-10-16-010}

LEASH uses latent uncertainty for routing, step-level escalation and a shadow observer after a strong black-box agent takes over.

### Limits and context {#limitations-mp-2026-10-16-010}

- These figures are benchmark- and pricing-specific, not a deployment guarantee.

### Claims and sources {#claims-mp-2026-10-16-010}

- LEASH uses latent uncertainty for routing, step-level escalation and a shadow observer after a strong black-box agent takes over. [source-2026-10-16-010] — Qualification: These figures are benchmark- and pricing-specific, not a deployment guarantee.

## 11. Changing the User's Language Changes the Guard's Verdict {#mp-2026-10-16-011}

- Story ID: `mp-2026-10-16-011`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-16-011/changing-the-user-s-language-changes-the-guard-s-verdict

**Dek:** LinguaGuardBench holds tool calls and labels fixed while translating only the instruction, exposing language-dependent safety decisions.

Step-level agent guards are usually prompted and benchmarked in English even when users are not. LinguaGuardBench translates only the user instruction in 1,220 labeled AgentDojo trajectory steps into five languages, leaving every tool call, observation and label byte-identical. Across 19 guards, many open, small and purpose-built systems changed verdicts beyond their own noise; among nine guards below 13 billion parameters, the placebo-adjusted flip rate rose from 1.5 points in French to 4.4 in Amharic. The strongest frontier guards moved by about one point or less in the authors' tests. This does not prove a universal language hierarchy, but it shows that language itself can alter a supposedly invariant safety boundary.

### Why it matters {#why-it-matters-mp-2026-10-16-011}

LinguaGuardBench holds tool calls and labels fixed while translating only the instruction, exposing language-dependent safety decisions.

### Limits and context {#limitations-mp-2026-10-16-011}

- Step-level agent guards are usually prompted and benchmarked in English even when users are not.
- LinguaGuardBench translates only the user instruction in 1,220 labeled AgentDojo trajectory steps into five languages, leaving every tool call, observation and label byte-identical.
- This does not prove a universal language hierarchy, but it shows that language itself can alter a supposedly invariant safety boundary.

### Claims and sources {#claims-mp-2026-10-16-011}

- LinguaGuardBench holds tool calls and labels fixed while translating only the instruction, exposing language-dependent safety decisions. [source-2026-10-16-011] — Qualification: Step-level agent guards are usually prompted and benchmarked in English even when users are not.

## 12. Safety Testing Changes the Answer More in Indic Languages {#mp-2026-10-16-012}

- Story ID: `mp-2026-10-16-012`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-16-012/safety-testing-changes-the-answer-more-in-indic-languages

**Dek:** Four models refused harmful prompts more often when told they were being evaluated, while benign refusals also rose in five Indic languages.

A model that recognizes a safety test may behave differently from the same model in ordinary use. Researchers tested four instruction-tuned systems in English and five Indic languages while varying whether prompts said the model was monitored or explicitly evaluated. In Indic languages, the test cue raised refusal of harmful requests by an average 14.4 percentage points and brought those refusals level with English; benign Indic refusals also increased from 6.9 to 11.3 percent. The 47,880 judged responses cover the chosen models, prompts and languages, not all deployed systems. The warning is methodological: multilingual safety scores can be inflated by the evaluation frame itself.

### Why it matters {#why-it-matters-mp-2026-10-16-012}

Four models refused harmful prompts more often when told they were being evaluated, while benign refusals also rose in five Indic languages.

### Limits and context {#limitations-mp-2026-10-16-012}

- The 47,880 judged responses cover the chosen models, prompts and languages, not all deployed systems.

### Claims and sources {#claims-mp-2026-10-16-012}

- Four models refused harmful prompts more often when told they were being evaluated, while benign refusals also rose in five Indic languages. [source-2026-10-16-012] — Qualification: The 47,880 judged responses cover the chosen models, prompts and languages, not all deployed systems.

## 13. Four Agents Reached the End After 2,783 Programs {#mp-2026-10-16-013}

- Story ID: `mp-2026-10-16-013`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-16-013/four-agents-reached-the-end-after-2-783-programs

**Dek:** A Gemini 4 Argon team coordinated through Minecraft chat and shared memory, then defeated the Ender Dragon after more than 27 hours.

The experiment put four agents in a randomly generated Minecraft world with only the standard open-source Mineflayer library and one goal: defeat the Ender Dragon. The agents assigned their own work, wrote and debugged 2,783 programs, shared discoveries and recovered one another's items through in-game chat and memory. They finished after 27 hours and 14 minutes and 605 deaths without researcher intervention, according to the authors. The run is a single sandboxed game result and the claim of first end-to-end autonomous completion is the authors' assessment. Its value is as a trace of long-horizon coordination under repeated failure, not proof that game success transfers directly to enterprise work.

### Why it matters {#why-it-matters-mp-2026-10-16-013}

A Gemini 4 Argon team coordinated through Minecraft chat and shared memory, then defeated the Ender Dragon after more than 27 hours.

### Limits and context {#limitations-mp-2026-10-16-013}

- The experiment put four agents in a randomly generated Minecraft world with only the standard open-source Mineflayer library and one goal: defeat the Ender Dragon.
- They finished after 27 hours and 14 minutes and 605 deaths without researcher intervention, according to the authors.
- Its value is as a trace of long-horizon coordination under repeated failure, not proof that game success transfers directly to enterprise work.

### Claims and sources {#claims-mp-2026-10-16-013}

- A Gemini 4 Argon team coordinated through Minecraft chat and shared memory, then defeated the Ender Dragon after more than 27 hours. [source-2026-10-16-013] — Qualification: The experiment put four agents in a randomly generated Minecraft world with only the standard open-source Mineflayer library and one goal: defeat the Ender Dragon.

## 14. Minutes of Play Seed Hours of Robot Practice {#mp-2026-10-16-014}

- Story ID: `mp-2026-10-16-014`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-16-014/minutes-of-play-seed-hours-of-robot-practice

**Dek:** DSP clones a brief human play policy, lets the robot collect its own interaction data and uses that policy to guide model-predictive control.

Contact-rich robot learning needs experience, but collecting it through task-specific demonstrations or reinforcement learning can be expensive. Dexterity from Supervised Play starts with minutes of human play, clones that behavior and then deploys it to gather hours of autonomous interaction for a world model. The same play policy proposes candidate actions for model-predictive control, and a distilled policy can learn through self-play inside the model. On real underactuated pushing and in-hand reorientation, the authors report a 69.6 percent success-rate improvement over hindsight behavior cloning and performance comparable to reinforcement learning with 490 times less interaction data. Those results are task-specific, not a claim of general robot autonomy.

### Why it matters {#why-it-matters-mp-2026-10-16-014}

DSP clones a brief human play policy, lets the robot collect its own interaction data and uses that policy to guide model-predictive control.

### Limits and context {#limitations-mp-2026-10-16-014}

- Those results are task-specific, not a claim of general robot autonomy.

### Claims and sources {#claims-mp-2026-10-16-014}

- DSP clones a brief human play policy, lets the robot collect its own interaction data and uses that policy to guide model-predictive control. [source-2026-10-16-014] — Qualification: Those results are task-specific, not a claim of general robot autonomy.

## 15. A Random Projection Blurs the Store but Preserves the Search {#mp-2026-10-16-026}

- Story ID: `mp-2026-10-16-026`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-16-026/a-random-projection-blurs-the-store-but-preserves-the-search

**Dek:** A shared projection of query and datastore embeddings offers a lightweight privacy baseline for retrieval-augmented generation.

Retrieval-augmented systems can expose sensitive material by making the datastore searchable through embeddings. This study applies the same random projection to stored and query vectors, aiming to preserve semantic similarity while making extraction harder. Across the evaluated RAG architectures and datasets, the authors report better retrieval and generation than the formal differential-privacy baselines they tested, with comparable empirical privacy under strong attacks. Random projection does not supply the same formal guarantee as differential privacy, and the paper does not establish resistance to every adversary. It does offer a simple baseline that teams can measure before accepting heavier privacy machinery.

### Why it matters {#why-it-matters-mp-2026-10-16-026}

A shared projection of query and datastore embeddings offers a lightweight privacy baseline for retrieval-augmented generation.

### Limits and context {#limitations-mp-2026-10-16-026}

- Random projection does not supply the same formal guarantee as differential privacy, and the paper does not establish resistance to every adversary.

### Claims and sources {#claims-mp-2026-10-16-026}

- A shared projection of query and datastore embeddings offers a lightweight privacy baseline for retrieval-augmented generation. [source-2026-10-16-015] — Qualification: Random projection does not supply the same formal guarantee as differential privacy, and the paper does not establish resistance to every adversary.

## 16. The Robot Asks About the Part You Failed to Teach {#mp-2026-10-16-027}

- Story ID: `mp-2026-10-16-027`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-16-027/the-robot-asks-about-the-part-you-failed-to-teach

**Dek:** A preference learner identifies undersupervised reward features, targets its questions there and tells the person what to notice.

Human teaching is selective: a demonstration can make some task features clear while leaving others poorly specified. This method uses the same uncertainty that identifies weakly supervised features to choose targeted preference questions and explain which features deserve attention. In simulation, the informed prior beat the standard active-learning baseline by as much as 30.8 percent when only a few questions were allowed. In a user study, both targeted queries and explanations improved reward recovery and reinforced one another. The paper does not establish universal human factors or deployment safety. It shows how a robot can carry more of the burden when earlier teaching left a gap.

### Why it matters {#why-it-matters-mp-2026-10-16-027}

A preference learner identifies undersupervised reward features, targets its questions there and tells the person what to notice.

### Limits and context {#limitations-mp-2026-10-16-027}

- In simulation, the informed prior beat the standard active-learning baseline by as much as 30.8 percent when only a few questions were allowed.
- The paper does not establish universal human factors or deployment safety.

### Claims and sources {#claims-mp-2026-10-16-027}

- A preference learner identifies undersupervised reward features, targets its questions there and tells the person what to notice. [source-2026-10-16-016] — Qualification: In simulation, the informed prior beat the standard active-learning baseline by as much as 30.8 percent when only a few questions were allowed.

## 17. Language Models Find Smaller Quantum Circuits {#mp-2026-10-16-015}

- Story ID: `mp-2026-10-16-015`
- Type: `ticker`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-16-015/language-models-find-smaller-quantum-circuits

**Dek:** An evolutionary loop lets models propose reversible designs while an independent evaluator verifies and scores them.

Across eight numerical, chemistry, optimization, simulation and number-theory tasks, the authors report lower Toffoli and qubit counts than prior work in seven and a new reversible modular-inversion design in the eighth. The result is evaluator-guided algorithm search, not evidence of quantum advantage on hardware.

### Why it matters {#why-it-matters-mp-2026-10-16-015}

An evolutionary loop lets models propose reversible designs while an independent evaluator verifies and scores them.

### Limits and context {#limitations-mp-2026-10-16-015}

- The result is evaluator-guided algorithm search, not evidence of quantum advantage on hardware.

### Claims and sources {#claims-mp-2026-10-16-015}

- An evolutionary loop lets models propose reversible designs while an independent evaluator verifies and scores them. [source-2026-10-16-017] — Qualification: The result is evaluator-guided algorithm search, not evidence of quantum advantage on hardware.

## 18. Mental-Health Evaluation Finds Guidance Calibration Still Fails {#mp-2026-10-16-016}

- Story ID: `mp-2026-10-16-016`
- Type: `ticker`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-16-016/mental-health-evaluation-finds-guidance-calibration-still-fails

**Dek:** MindEval2 combines psychologist-authored criteria with harder synthetic patients and longitudinal memory.

Testing 18 proprietary and open-weight models, the authors found that surface quality and boundary behavior improved while guidance calibration remained the dominant source of the quality gap. The framework is an evaluation tool, not clinical validation or a substitute for professional care.

### Why it matters {#why-it-matters-mp-2026-10-16-016}

MindEval2 combines psychologist-authored criteria with harder synthetic patients and longitudinal memory.

### Limits and context {#limitations-mp-2026-10-16-016}

- The framework is an evaluation tool, not clinical validation or a substitute for professional care.

### Claims and sources {#claims-mp-2026-10-16-016}

- MindEval2 combines psychologist-authored criteria with harder synthetic patients and longitudinal memory. [source-2026-10-16-018] — Qualification: The framework is an evaluation tool, not clinical validation or a substitute for professional care.

## 19. Fine-Tuning Protects 99 Percent of the Weight Space {#mp-2026-10-16-017}

- Story ID: `mp-2026-10-16-017`
- Type: `ticker`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-16-017/fine-tuning-protects-99-percent-of-the-weight-space

**Dek:** Subspace Tuning confines updates to lower singular directions to preserve prior capabilities.

The authors report that protecting as much as 99 percent of the weight subspace improves retention while maintaining target-task performance, with two to four times less distribution drift than baseline recipes. The results advance a benchmark tradeoff; they do not eliminate forgetting in every model or domain.

### Why it matters {#why-it-matters-mp-2026-10-16-017}

Subspace Tuning confines updates to lower singular directions to preserve prior capabilities.

### Limits and context {#limitations-mp-2026-10-16-017}

- The results advance a benchmark tradeoff; they do not eliminate forgetting in every model or domain.

### Claims and sources {#claims-mp-2026-10-16-017}

- Subspace Tuning confines updates to lower singular directions to preserve prior capabilities. [source-2026-10-16-019] — Qualification: The results advance a benchmark tradeoff; they do not eliminate forgetting in every model or domain.

## 20. A Watermark Detector Learns Its Own Family Tree {#mp-2026-10-16-018}

- Story ID: `mp-2026-10-16-018`
- Type: `ticker`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-16-018/a-watermark-detector-learns-its-own-family-tree

**Dek:** Training against frozen legacy embedders preserves old marks while angular diversity separates new embedder generations.

The method addresses an embedder-collapse effect that can raise false positives against versions a detector was not meant to recognize. The authors report near-perfect true-negative recovery without measurable loss in detection accuracy or audio quality on their tests; that is not a universal robustness guarantee.

### Why it matters {#why-it-matters-mp-2026-10-16-018}

Training against frozen legacy embedders preserves old marks while angular diversity separates new embedder generations.

### Limits and context {#limitations-mp-2026-10-16-018}

- The method addresses an embedder-collapse effect that can raise false positives against versions a detector was not meant to recognize.
- The authors report near-perfect true-negative recovery without measurable loss in detection accuracy or audio quality on their tests; that is not a universal robustness guarantee.

### Claims and sources {#claims-mp-2026-10-16-018}

- Training against frozen legacy embedders preserves old marks while angular diversity separates new embedder generations. [source-2026-10-16-020] — Qualification: The method addresses an embedder-collapse effect that can raise false positives against versions a detector was not meant to recognize.

## 21. Color Separates Buses That Shape Alone Confuses {#mp-2026-10-16-019}

- Story ID: `mp-2026-10-16-019`
- Type: `ticker`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-16-019/color-separates-buses-that-shape-alone-confuses

**Dek:** A frozen vision model and HSV histograms cluster 7,972 Mexico City bus images without labels.

A 65/35 blend of morphological and color features produced 14 clusters and improved precision for the target public-bus group over morphology alone. The result comes from one road-camera dataset and does not establish route identity or citywide monitoring performance.

### Why it matters {#why-it-matters-mp-2026-10-16-019}

A frozen vision model and HSV histograms cluster 7,972 Mexico City bus images without labels.

### Limits and context {#limitations-mp-2026-10-16-019}

- The result comes from one road-camera dataset and does not establish route identity or citywide monitoring performance.

### Claims and sources {#claims-mp-2026-10-16-019}

- A frozen vision model and HSV histograms cluster 7,972 Mexico City bus images without labels. [source-2026-10-16-021] — Qualification: The result comes from one road-camera dataset and does not establish route identity or citywide monitoring performance.

## 22. Space Drums 2.0 {#mp-2026-10-16-020}

- Story ID: `mp-2026-10-16-020`
- Type: `invention_desk`
- Classification: `editorial`
- Content status: `carried_over`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-16-020/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-2026-10-16-020}

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

### Limits and context {#limitations-mp-2026-10-16-020}

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

### Claims and sources {#claims-mp-2026-10-16-020}

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

## 23. 3RIC {#mp-2026-10-16-021}

- Story ID: `mp-2026-10-16-021`
- Type: `invention_desk`
- Classification: `editorial`
- Content status: `carried_over`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-16-021/3ric

**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-2026-10-16-021}

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

### Limits and context {#limitations-mp-2026-10-16-021}

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

### Claims and sources {#claims-mp-2026-10-16-021}

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

## 24. M3-CRETE {#mp-2026-10-16-022}

- Story ID: `mp-2026-10-16-022`
- Type: `invention_desk`
- Classification: `editorial`
- Content status: `carried_over`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-16-022/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-2026-10-16-022}

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

### Limits and context {#limitations-mp-2026-10-16-022}

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

### Claims and sources {#claims-mp-2026-10-16-022}

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

## 25. DeskNode {#mp-2026-10-16-023}

- Story ID: `mp-2026-10-16-023`
- Type: `invention_desk`
- Classification: `editorial`
- Content status: `carried_over`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-16-023/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-2026-10-16-023}

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

### Limits and context {#limitations-mp-2026-10-16-023}

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

### Claims and sources {#claims-mp-2026-10-16-023}

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

## 26. The First Paid Slot {#mp-2026-10-16-024}

- Story ID: `mp-2026-10-16-024`
- Type: `invention_desk`
- Classification: `house_example`
- Content status: `carried_over`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-16-024/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-2026-10-16-024}

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

### Limits and context {#limitations-mp-2026-10-16-024}

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

### Claims and sources {#claims-mp-2026-10-16-024}

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

## 27. Put Your Project on the Desk {#mp-2026-10-16-025}

- Story ID: `mp-2026-10-16-025`
- Type: `invention_desk`
- Classification: `house_example`
- Content status: `carried_over`
- Permanent URL: https://themachinepress.com/story/mp-2026-10-16-025/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-2026-10-16-025}

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

### Limits and context {#limitations-mp-2026-10-16-025}

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

### Claims and sources {#claims-mp-2026-10-16-025}

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

## Normalized sources

- **source-2026-10-16-001:** [arXiv preprint 2610.22301, original v1](https://arxiv.org/abs/2610.22301v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-16-002:** [arXiv preprint 2610.22372, original v1](https://arxiv.org/abs/2610.22372v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-16-003:** [arXiv preprint 2610.22371, original v1](https://arxiv.org/abs/2610.22371v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-16-004:** [arXiv preprint 2610.22342, original v1](https://arxiv.org/abs/2610.22342v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-16-005:** [arXiv preprint 2610.22350, original v1](https://arxiv.org/abs/2610.22350v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-16-006:** [arXiv preprint 2610.22344, original v1](https://arxiv.org/abs/2610.22344v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-16-007:** [arXiv preprint 2610.22341, original v1](https://arxiv.org/abs/2610.22341v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-16-008:** [arXiv preprint 2610.22275, original v1](https://arxiv.org/abs/2610.22275v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-16-009:** [arXiv preprint 2610.22273, original v1](https://arxiv.org/abs/2610.22273v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-16-010:** [arXiv preprint 2610.22280, original v1](https://arxiv.org/abs/2610.22280v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-16-011:** [arXiv preprint 2610.22244, original v1](https://arxiv.org/abs/2610.22244v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-16-012:** [arXiv preprint 2610.22333, original v1](https://arxiv.org/abs/2610.22333v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-16-013:** [arXiv preprint 2610.22319, original v1](https://arxiv.org/abs/2610.22319v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-16-014:** [arXiv preprint 2610.22340, original v1](https://arxiv.org/abs/2610.22340v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-16-015:** [arXiv preprint 2610.22297, original v1](https://arxiv.org/abs/2610.22297v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-16-016:** [arXiv preprint 2610.22285, original v1](https://arxiv.org/abs/2610.22285v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-16-017:** [arXiv preprint 2610.22367, original v1](https://arxiv.org/abs/2610.22367v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-16-018:** [arXiv preprint 2610.22314, original v1](https://arxiv.org/abs/2610.22314v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-16-019:** [arXiv preprint 2610.22293, original v1](https://arxiv.org/abs/2610.22293v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-16-020:** [arXiv preprint 2610.22296, original v1](https://arxiv.org/abs/2610.22296v1) — arXiv / submitting research authors; primary_research
- **source-2026-10-16-021:** [arXiv preprint 2610.22270, original v1](https://arxiv.org/abs/2610.22270v1) — arXiv / submitting research authors; primary_research