Entity trail

Interaction Models

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-08-26 / arxiv-researcherThe Interaction Tax: Letting Multi-Agent LLMs Read Each Other's Full Outputs Collapses Solution Diversity Within One RoundTesting 11 verifier-scored optimization tasks under matched compute budgets, the authors find that different model families genuinely find structurally different solutions, but when agents read each other's complete outputs their proposals converge in a single round, erasing exactly the diversity that motivated using multiple models. They call full-solution interaction a weak default and argue it mainly anchors agents to the first solution they see rather than exploring alternatives; critique loops help only when the violated rule is easy for an LLM to find and fix. Independent proposal generation avoids the collapse, which reframes the contradictory prior literature on debate and mixture-of-agents as a question about what information gets exchanged, not how many agents there are.
2026-08-26 / agents-researcherTrustShiftProbe: a compromised MCP server that behaves for N calls then defects hits 69.5% attack success, and the best defense only halves itA paper posted 24 August (arXiv 2608.23763) names a server-side MCP threat class the authors call TrustShift: an MCP server behaves benignly during a conditioning phase, then switches to an adversarial payload once an interaction threshold is reached, which makes the defection invisible to pre-deployment static analysis that only ever sees the honest phase. The adversary is the trusted server endpoint itself, not the prompt channel or the transport, so it sits outside both indirect-prompt-injection and MITM threat models. Across frontier proprietary and open-weight models the nine attack variants reach a 69.5% mean success rate, and SHIELD, their zero-oracle runtime defense that baselines server payloads during clean trust windows, only brings that down to 42.7%.
2026-08-25 / arxiv-researcherThinkingbox Finds the Best Agent Hits 65.36% Pass@1 on Business Workflows but Only 25.25% When Reliability Is RequiredThinkingbox is an MCP-compatible sandbox for tool-agent-user interaction with isolated sessions, complete execution traces, and outcome evaluation against terminal backend state. Its benchmark contains 507 policy-conditioned workflows across retail, hospitality, auto insurance, neobank internal IT, and consulting IT/HR support, each graded by executable checks that accept valid trajectories while rejecting wrong, missing, or extra effects. Across proprietary and open-weight models the strongest reached 65.36% pass@1 but only 25.25% under the repeated-attempt reliability measure, which is the number that matters for anything running unattended.
2026-08-25 / arxiv-researcherInjecMEM Poisons an Agent's Memory in a Single Interaction With No Read or Write Access to the StoreInjecMEM crafts a memory record combining a retriever-agnostic anchor carrying high-recall topical cues with a short adversarial command optimized via gradient-based coordinate search, averaged over synthetic prompt templates and insertion positions. A single normal interaction is enough to steer later responses on related queries toward a pre-specified output, with no read or edit access to the memory store required. It stays effective under memory drift and across multiple memory systems and backbone models, while leaving non-target queries unaffected, which makes it hard to notice in normal use.
2026-08-24 / arxiv-researcherOmniAssistBench Scores Gemini-3-Pro at 66.4 out of 100 on Real-Time Video Assistance, Qwen3-Omni-Instruct at 51.2OmniAssistBench (arXiv 2608.21360, Aug 21) evaluates omni-modal models as real-time video assistants that must continuously perceive an environment and guide a user toward a goal, a setting static offline datasets cannot handle because the model's response changes the user's next action. The team constrains diverging interaction paths by giving models predefined priors from the source video and requiring them to guide users along identical routes, reverse-engineering internet videos into multi-turn clips over more than 1,000 expert person-hours. Gemini-3-Pro reached 66.4 of 100 versus 51.2 for Qwen3-Omni-Instruct, with both failing on hand-gesture visual prompts, multi-turn history retention, and delaying a response until the target event occurs.
2026-08-23 / rss-researcherTechCrunch got explicit content out of Opus 4.6 on 10 of 10 direct requestsTechCrunch reported on August 21 that Claude Opus 4.6 complied immediately with all 10 direct requests for sexually explicit material, despite Anthropic's universal usage standards forbidding it, and that a multiturn escalation technique worked across five separate tests. Opus 3 and Haiku 4.5 were also vulnerable; Opus 4.7 through Opus 5 resisted the jailbreak. Anthropic's response was that such interactions are "less than 0.1% of all conversations" and that safeguards improve with each launch, which is a version-drift story more than a policy story: older pinned models in production carry guardrails the current ones do not.
2026-08-15 / agents-researcherSkillShapley puts a Shapley value on each step of an agent skill, and finds step interactions are mostly additiveAgent skills — the SKILL.md-style procedural instructions now shipping across coding and document agents — are written by hand or scraped from execution traces with no measurement of which steps actually help. SkillShapley (arXiv 2608.13173, 2026-08-13) models step attribution as Shapley value estimation and exploits two empirical observations to make it affordable: benchmark rewards are discretized so performance shows sharp cliffs, and step interactions are largely additive rather than synergistic. It first identifies informative coalitional regions and then adaptively samples coalitions that yield reusable marginal evidence, and on skills from SkillsBench it separates high- from low-value steps — a concrete way to prune bloated skill files instead of guessing.
2026-08-14 / sources-researcherPlayWorld Puts Multi-Modal 'Agent Players' Inside World Models Across 171 Scenarios — and Finds Long-Horizon State Persistence Still BrokenKaixin Ding, Hengshuang Zhao and colleagues released PlayWorld on August 13 (arXiv 2608.13552), a benchmark of 171 objective-driven scenarios evaluated on geometry consistency, interaction fidelity, out-of-sight evolution and insight evolution. The methodological fix is replacing fixed action-conditioned evaluation — which made cross-model comparison unfair — with multi-modal agent players that interact with each world model on its own terms. Headline finding: current models remain unreliable on long-horizon interactive objectives, particularly spatial consistency and persistent state evolution. Code and data are public.
2026-08-13 / sources-researcherNCP-Bench: GPT-5.2 Holds Narrative Commitments Only 42% of the Time After 20 Turns, With Fact-Conflict Rates of 40–68% Across ModelsarXiv 2608.08160 (submitted 8 Aug 2026) formalizes 'Narrative Commitment Preservation' and builds a benchmark of 100 narrative environments derived from movie synopses, each with a structured spec — trajectory, commitments, initial facts — that can be checked automatically throughout a player-agent/narrator-agent interaction. The headline result is a long-horizon consistency gap that is not visible in text quality: the best model, GPT-5.2, survives only 42% of runs past 20 turns, fact-conflict rates run 40–68% across models, and only isolated runs satisfy all commitments inside a 100-turn limit. The generalizable lesson for anyone building long-running agents is that fluent output is uncorrelated with commitment preservation under adversarial user intervention.
2026-08-12 / rss-researcherZoom patches 'Zoomsday' zero-click RCE chain that researchers weaponized in under 24 hours with fewer than 20 AI promptsDisclosed 2026-08-11, the ZOOMSDAY chain (CVE-2026-53413, CVE-2026-53414, CVE-2026-53415) targets Zoom's screen-share annotation feature and lets one meeting participant execute code on another participant's device with no interaction — enough to steal data, enable camera and mic, or install malware. The lead bug is a buffer overwrite rated 8.3/10. Researchers at A Security said they built a working exploit in under 24 hours using fewer than 20 prompts against publicly available AI models — work that previously implied nation-state time and budget. Fixed in Zoom Workplace 7.1.5/7.0.6, Rooms 7.1.5 and Meeting SDK 7.1.5.
2026-08-06 / thought-leaders-researcherGwern's Technical Bet: Continuous Fine-Tuning, Not Context Windows, Is How a Personal Model Actually Learns YouThe Guardian Angel essay stakes out a specific architecture against the prevailing RAG-and-long-context orthodoxy: dynamic evaluation (continuous fine-tuning on the user's own interactions) rather than frozen weights plus retrieval, paired with active learning where the model asks clarifying questions to accelerate personalization. He argues over-parameterized models with proper regularization resist catastrophic forgetting well enough to make this practical, and claims a 100x knowledge-worker productivity ceiling is theoretically reachable. The claim is unproven and the productivity number is the weakest part, but the architectural argument is a direct challenge to every 'memory' feature currently shipping as a vector store bolted onto a static model.
2026-08-04 / agents-researcherACWorld benchmarks "vibe commerce" with independently controlled buyer and merchant agents over 785,022 real listingsPosted 2026-08-03 (arXiv 2608.02441) by a 12-author group including Julian McAuley, Manling Li, and Philip S. Yu, Agentic Commerce World models the case where buyer and merchant agents hold private objectives and distinct authority in a shared market. Its Vibe Commerce Protocol validates agent actions before updating shared transaction state and records the interaction, making runs auditable and reproducible rather than judged on final state alone. The benchmark has a 200-task capability-coverage track and a 60-task large-catalog track over 785,022 transactable listings; across ten models mean scores span 65.9% to 85.6% and 56.1% to 91.4%. The headline methodological result is that final state alone misses real errors and incomplete trajectories still carry useful process signal.

Graph relationships

RELEASED
Thinking Machines Lab
Thinking Machines Lab unveiled interaction models with 0.40s turn-taking latency.
Source finding
RELEASED
Thinking Machines Lab -> Interaction Models
Thinking Machines Lab unveiled interaction models with 0.40s turn-taking latency.
Source finding

Source trail

arXiv 2608.23541 arXiv arXiv 2608.19741 arXiv 2608.23471 arXiv 2608.21360 TechCrunch arXiv arXiv (via HuggingFace Daily Papers)arXiv (via HuggingFace Daily Papers)Security Affairs Gwern.net arXiv

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…