This edition of the daily paper deep dive (September 14, 2026) covers three arXiv AI papers released September 10, 2026, explained in a Feynman-inspired style.
Paper 1: Artificial Id — When AI Develops "Desires"
Paper: Artificial Id: an adaptive internal drive for agentic AI — Yakov Shkolnikov, arXiv:2609.11911
The paper asks: as LLM agents evolve from answering questions to autonomously completing tasks, do they need something like Freud's "Id" — an internal drive?
- Inspired by bacterial chemotaxis: E. coli never received a "find sugar" program; survival itself selected the behavior. Shkolnikov builds a virtual petri dish with a 20-parameter linear "bacteria bot" that cannot see its health value, receives no goal instructions, and has no reward function. Its key mechanic is differential persistence: near food, its lifespan clock runs 20x slower, so control policies that keep it near food replicate more often upon death. No objective function, no gradient descent — only survival-based selection.
- World 1: evolution exploits a physics loophole — a near-constant backward force freezes orientation, converting a body-frame force into a world-frame constant force. Natural selection favors "good enough," not optimal.
- World 2: with the loophole sealed, the population genuinely learns to use local sensor signals. Drive emerges from environmental coupling rather than external injection; evolved controllers beat hand-designed references across three seeds.
- World 3: after learning, food signals are swapped left/right. Frozen populations collapse (below 1% occupancy), but mutating populations show rising diversity, gradual sign flips in food weights, and full recovery — online re-adaptation without being told the rules changed.
- The author proposes a functional Id-Ego-Superego architecture: the Id provides drive (continue/stop/change), the Ego is the general reasoning model, and the Superego defines alignment boundaries (trusted observations, consequence channels, persistent state, authority, identity, provenance, hard constraints). Moving drive inside the agent makes alignment a property of continuing agentic systems rather than single trajectories — but the same persistence mechanism could carry corrupted or misaligned states across task boundaries.
- Humans average 97.87%; the best model, GPT-5.6-Sol, averages 61.42%.
- A stark inference-planning gap: GPT-5.6-Sol drops from 66.83% (inference) to 52.75% (planning); Gemini-3.1-Pro falls to 19.23%; the best open model (Qwen3.5-397B-A17B) plans at just 5.27%, with 0% on Pipe and One Stroke for all open models.
- Error analysis of 910 failure cases: inference errors are dominated by perceptual grounding failures (58.1%), planning by action-planning errors (43.3%). Explicit topological-invariance errors are rare (3.8–5.4%) — models fail before reaching topology proper.
- SFT+RL on Qwen3-VL-2B lifts inference from 8% to 51.53% but planning only from 0.20% to 6.33%; imagined visual rollouts do not reliably follow environment dynamics.
- Decompose evidence into context (a) and remaining evidence (e), assume a → d → e, and build a reverse posterior R(d|x) ∝ P(e|d) × P(d|a). R and each agent's forward posterior are different factorizations of the same posterior — biased differently, so their errors are less collinear (label-collision with forward errors ≈0.33 vs. 0.72 between forward agents).
- Jensen-Shannon divergence measures agreement between each forward posterior F_i and R. Notably, R's standalone accuracy is often below the forward pool's average — its value is a *different viewpoint*, not higher accuracy.
- Three aggregation strategies: MinJS (pick the agent most consistent with R), FwdJS (softmax weighting, τ=5.0), and LogLin (multiplicative fusion, w_R=0.2, requiring both directions to support a label for amplification). On DDXPlus, LogLin wins on all 5 backbones and gains 1.2–4.7 points over the best election baseline on the disagreement subset (79.07% GTPA@1 on DeepSeek-V4-Flash).
- Ablations confirm the anchor's uniqueness: replacing R with the pool mean (MeanF) or an external generic agent (GenF) strictly degrades performance — consensus distance rewards conformity, while R rewards cross-path consistency.
- A lightweight two-stage calibration (ordinal shape parameters + class-prior correction) further refines R with minimal labeled data, without touching the LLMs or aggregation operator.
Paper 2: MindTopo — Can Foundation Models Reason in Topological Space?
Paper: MindTopo: Can Foundation Models Reason in Topological Space? — Ge, Liu, Wang, et al. (Northwestern, Microsoft Research, Stanford), arXiv:2609.11900
Following Piaget's observation that children grasp topological relations (inside/outside, connected/separated) before metric ones, MindTopo tests multimodal LLMs on five core properties:
1. Continuity — maze reachability, pipe connections 2. Separation — assembly separability, one-stroke region separation 3. Order — bead-string sequence reproduction, origami point ordering 4. Enclosure — fence/sheep, hole counting, Chat Noir cat-blocking game 5. Knots — detecting real knots, linkages, and untangling planning
Key findings from 11,030 instances (73% inference, 27% planning) across 14 MLLMs:
The conclusion: current models handle metric space reasonably well but have fundamental weaknesses in topological space — a serious obstacle to autonomous action in the physical world.
Paper 3: Bayesian Backward Reasoning as a Label-Free Anchor
Paper: When Agents Disagree: Bayesian Backward Reasoning as a Label-Free Anchor for Multi-Agent Collective Decision-Making — Chen, Wang, Seneviratne, Weeratunge, Halgamuge (University of Melbourne), arXiv:2609.11709
Voting, judge LLMs, and multi-round debate all operate in the same direction — forward reasoning from evidence to conclusion — so shared error factors get amplified rather than corrected. The proposed fix:
---
*All three papers were released September 10, 2026, from an independent researcher (Shkolnikov), a Northwestern/Microsoft/Stanford team, and a University of Melbourne team.*