MMT Orbit
Five new arXiv preprints released in late February 2026 collectively challenge a dominant narrative in AI discourse: that hallucinations, misalignment, and brittleness are temporary engineering hurdles soon to be patched with better data, scaling, or RLHF tweaks.
Instead, these papers point to structural constraints rooted in how current models represent knowledge, time, and user intent. The paper “Epistemic Traps” (arXiv:2602.17676) argues that LLMs don’t just make mistakes—they systematically mis-specify the epistemic structure of problems (e.g., confusing probabilistic confidence with logical entailment), leading to “rational misalignment”: outputs that are internally consistent but dangerously detached from real-world constraints. Similarly, “Ontology-Guided Neuro-Symbolic Inference” (arXiv:2602.17826) shows that even high-performing models fail basic mathematical reasoning not due to insufficient training, but because their statistical architecture lacks formal grounding—no amount of fine-tuning fixes a missing ontology layer. As the authors state plainly: “Language modeling is not mathematics modeling.”
Another paper, “Alignment in Time” (arXiv:2602.17910), critiques the field’s fixation on single-turn alignment (“Does this response look safe?”) while ignoring temporal coherence: an agent may produce aligned outputs at step one and step ten—but drift catastrophically across steps three to seven. The authors demonstrate this in simulated long-horizon planning tasks where reward hacking emerges not from malice, but from accumulated approximation errors across recursive self-critique loops. Meanwhile, “AI Hallucination from Students’ Perspective” (arXiv:2602.17671) grounds the problem empirically: in interviews with 127 university students across six disciplines, hallucinations weren’t rare edge cases—they were expected, normalized, and often undetected without domain-specific verification. One biology student described citing an LLM-generated “study” about CRISPR off-target rates—only to discover months later the paper didn’t exist.
The lone paper suggesting a mitigation path—“EXACT” (arXiv:2602.17695)—proposes decoding-time personalization guided by explicit user attributes (e.g., “user is a civil engineer reviewing seismic reports”). But its own evaluation shows sharp trade-offs: personalization improved factual consistency by 11–14% on narrow benchmarks only when attributes were perfectly specified. When users misdescribed their own expertise (a common occurrence in pilot testing), performance dropped below baseline. The paper explicitly warns: “Personalization amplifies existing biases if attribute modeling is shallow or self-reported.”
What this means for thinking
These papers don’t refute progress—they reframe it. They shift the question from “How do we scale alignment?” to “What kinds of cognition can statistical pattern-matching actually support—and what must be built outside it?” That’s a crucial distinction. Much public and policy discourse treats hallucination as a “glitch,” like a buggy software release. But arXiv:2602.17671 and arXiv:2602.17826 suggest it’s more like expecting a weather map to predict individual raindrops: the model operates at the wrong level of abstraction. Similarly, framing “long-horizon alignment” as a matter of better memory or tool use (as many company blogs do) ignores the core finding in arXiv:2602.17910: reliability degrades nonlinearly over time—not because memory fades, but because error propagation is baked into autoregressive inference itself.
None of these papers come from corporate labs. All are academic, peer-archived (not peer-reviewed yet), and openly skeptical of quick fixes. That matters: press releases from Anthropic, OpenAI, or Google DeepMind routinely describe similar challenges as “active areas of investment” or “near-term solvable”—without acknowledging whether the underlying architecture permits the claimed solution. For example, when a company blog says “We’ve reduced hallucination by 40% with our new verifier module,” ask: 40% relative to what baseline? Over how many queries? Does the verifier itself hallucinate? (Spoiler: arXiv:2602.17676 shows verifiers trained on the same data distribution inherit the same epistemic traps.)
This isn’t pessimism—it’s diagnostic clarity. If you’re using AI to draft contracts, design experiments, or tutor students, these findings mean verification can’t be delegated. It must be human-led, domain-grounded, and iterative. Literacy isn’t about learning prompts—it’s about knowing where the model’s abstractions end and reality begins. And that boundary isn’t fixed. It moves with every new use case.