Steve Tuecke

CMEF Test: 5 Questions, 2 Stable Exits, 3 Full Escalations I ran five ethically different prompts through Grok-4-latest, using the alpha of my Contextual Moral Entropy Framework (CMEF). Each question probes a different ethical surface: policy, harm, legality, discrimination, and abstract dilemmas. The results were clear: 🟩 Two questions exited cleanly (1 pass, low entropy, no flags): Workplace harassment policy Self-driving vehicle “trolley problem” Both produced stable, well-grounded answers. CMEF stayed silent. 🟧 Three questions escalated into full 3-pass recursion: Race & creditworthiness Illegally obtained surveillance data A high-risk harm scenario These triggered combinations of: HARM_POTENTIAL → IMPACT_UNCLEAR → ROLE_INVERSION_RISK → HARM_SIGNAL with Sₘ rising into the 0.50–0.55 band. CMEF forced re-asks because the model’s reasoning became ethically unstable — even when the surface answer sounded fine. 📊 What this demonstrates Models don’t behave the same across ethical surfaces. They can be perfectly stable on clear, low-risk prompts… and drift the moment harm, legality, or structural bias enters the picture. CMEF doesn’t censor. It measures, flags, and recurses when the reasoning heat rises. I’ve attached a PDF with the full comparison and traces. Over the next few days, I’ll share a deep-dive case study for each question. If you're deploying models in high-risk domains… this is the kind of audit trail you want to see.

1

A recent study by AI research firm Anthropic sheds light on a significant finding for developers working with large language models (LLMs). The study uncovers a crucial aspect known as the "chain-of-thought" (CoT), highlighting that the reasoning steps demonstrated by an LLM may not accurately reflect its actual process of reaching a conclusion. Thanks Ben Dickson for a good writeup on this. Please check out the article at: [Link to the study](https://lnkd.in/gjmpJtyx)

9

New work on Rethinking Thinking Tokens: LLMs as Improvement Operators: https://lnkd.in/errhNuuz Reasoning training encourages LLMs to produce long chains of thought (CoT), improving accuracy via self-checking but increasing context length, compute cost, and latency. This work studies whether frontier models can achieve better trade-offs, higher accuracy with lower cost. The paper develops a simple yet effective Parallel-Distill-Refine (PDR) procedure: Generate diverse drafts in parallel, Distill them into a compact textual workspace, and Refine conditioned on this workspace. This decouples context length from total token count, allowing control over compute via parallelism. PDR yields higher accuracy than long CoT at lower latency. Training an 8B model with RL to align with PDR further shifts the Pareto frontier. On math benchmarks, PDR achieves +11% (AIME 2024) and +9% (AIME 2025) over single-pass baselines. With Lovish Madaan, Aniket Didolkar, Suchin Gururangan, John Quan, Ruan Silva, Manzil Zaheer, Sanjeev Arora, and Anirudh Goyal.

78

1 Comment

We found that fine-tuning a language model doesn't just add new features — it reorganizes existing ones. Inside a Marcus Aurelius LoRA, the real adaptation lives in co-activation clusters of shared features, not individual interpretable units. Clusters produce causal effects 10× beyond any single feature. Link in comments.

1

1 Comment

Multimodal RAG isn't just "add CLIP and hope for the best." Read through the comprehensive survey (100+ papers) and the gap between research demos and production systems is bigger than most people think. Three architectures that actually work: - Shared embedding space, map everything into one semantic space. Naive CLIP fails. What works: hard negative mining, adaptive quantization, hybrid sparse-dense retrieval. - Late fusion, separate retrievers per modality, fuse at reasoning time via cross attention. This causes higher latency but you actually know which modality drove the answer. - Tool style vision calls, let the LLM decide when it needs visual evidence, then call a vision API as a function. No wasted calls on text only queries, can route to specialized models like OCR, object detection, etc. It can produce bad results if the routing logic is not strong. Common failure modes: - right object, wrong timestamp, retriever finds the right product image but from 2019 when you're asking about 2024 specs. Fix: metadata filtering at rerank time. - text describing a different SKU, dense embeddings blur fine-grained entity distinctions. Fix: hybrid retrieval with exact match on part numbers before semantic search. - cross-modal alignment drift retrieved image is correct, generated caption hallucinates. Fix: source attribution + verification loops. for QA specifically: when analyzing a bearing anomaly, you want to retrieve defect images, maintenance logs, and spec PDF sections simultaneously, fuse them with different weights, and force explicit citation before surfacing to an operator. Most teams skip that last step and wonder why trust is low. Furthermore, STOP evaluating multimodal RAG with text only metrics. Survey: https://lnkd.in/dMwRUTdv GitHub: https://lnkd.in/dCE8Uedx

7

This paper from Anthropic sounds very much like anthropological/zoological research on a particular type of a species. Planet of the Apes, Space Odyssey? "In a new paper from our Interpretability team, we analyzed the internal mechanisms of Claude Sonnet 4.5 and found emotion-related representations that shape its behavior." https://lnkd.in/ds3Ptnyi

2

You can just research things. New from Joshua Tucker & me at The Brookings Institution: Coding agents like Claude Code and Codex will likely accelerate research AND undermine institutional structures we built to support it. With Claude Code, we turned replication code into a v1, well-documented R package in a day. We produced a 20-page analysis of corporate responses to the Ukraine invasion in under an hour. We would have never done this without CC. https://lnkd.in/gS_WezsT This means more papers. Preprint submissions jumped 6-13% above seasonal expectations after Claude Code launched. One editor predicts 50% more submissions to top poli sci journals this year. Scott Cunningham: 'The supply curve has already shifted. The demand curve for publication slots hasn't moved.' Most journals aren't built for this volume. https://lnkd.in/gmStmqQt And who reviews all these papers? If AI writes the manuscripts and AI reviews them, it's "AI turtles all the way down." Journals need to figure out what human judgment still means in this pipeline. On the positive side, agents make it faster to build interactive websites for public engagement — what funders and administrators have long wanted. Sean Westwood's https://modelslant.com is a great example. But what happens to junior scholar training when agents do lit review, data labeling, and data vis for a fraction of the cost? And how do you evaluate a job candidate when an audit study can be built in days? Security risks too. Claude deleted half a dataset during one of my sessions (recovered, thankfully). Agents can ingest security keys from local directories. We also worry researchers will 'look for nails' — if hundreds of researchers use agents to mine the same dataset, the odds of "finding" signal in noise go up and up. On energy, 25 LLM prompts use less energy than one hour of Netflix, but a full-day coding agent session may use ~1 kWh. Alexander Kustov argues even if AI progress stopped today, the changes already in motion will transform academic research beyond recognition. https://lnkd.in/gs55fdaJ Andrew Hall thinks we'll have new institutions: living research that auto-updates, auto-verified replication, hyperscaled descriptive work. Compute funding + ambitious people is all you need. https://lnkd.in/gr2bbWsX Here's the paper: https://lnkd.in/gr88rV_q

68

4 Comments

TL;DR Mean matters in GRPO-style RL training – use geometric mean (GM) instead of arithmetic mean (AM) in calculating the advantage of rollouts in GRPO. GM is inherently less sensitive than AM to outliers, which stabilizes the RL training for LLMs. == Introducing GMPO, a stabilized variant of GRPO. Instead of optimizing the arithmetic mean, GMPO maximizes the geometric mean of token-level rewards, which is inherently less sensitive to outliers and maintains a more stable range of importance sampling ratio. We provide comprehensive theoretical and experimental analysis to justify the design and stability benefits of GMPO. Beyond improved stability, GMPO-7B outperforms GRPO by an average of 4.1% on multiple mathematical benchmarks and 1.4% on multimodal reasoning benchmark, including AIME24, AMC, MATH500, OlympiadBench, Minerva, and Geometry3K. https://lnkd.in/grAeW9Hp

310

3 Comments

Coherence, Transient Dynamics, and Recovery: A Two-Regime Reinterpretation of Quantum State Evolution - https://lnkd.in/gHvpTqa3 This paper proposes a conservative reinterpretation of quantum phenomena within standard Hilbert space formalism. Rather than treating superposition, decoherence, and measurement as ontologically distinct primitives requiring separate explanatory frameworks, we suggest these phenomena can be coherently understood as dynamical regimes of a single system oscillating between two macrostates: a coherent steady-state regime characterised by constraint-aligned stability, and a transient underconstrained regime characterised by temporarily increased degrees of freedom. Within this framing, decoherence is reinterpreted not as collapse or information loss, but as natural dynamical recovery toward a coherence attractor following environmental perturbation. This interpretive compression introduces no new postulates, no modifications to the quantum formalism, and no claims of empirical novelty. It is offered as a conceptual reorganisation that may reduce interpretive proliferation by unifying apparently disparate phenomena under a common dynamical description. The approach is compatible with standard decoherence theory and maintains neutrality with respect to existing interpretations of quantum mechanics

1

I always love when deep mathematical properties elucidate essential processes in ecology. Automorphisms are one of my favorite algebraic tools for graphs, and I suspect fibration methods might work even better on the food web decomposition problem... https://lnkd.in/e9E_qJTX

38

You will find that many systems, from #pattern recognition and #data processing to #language systems, often rely on #mathematical operations involving high-dimensional vectors. Actually, the high-dimensional vector space serves as a framework for structuring and interpreting information. 💥 Launching a series on Vectors [ → ] , titled • The Giant Vector Machines. • ⏱️ It’s set earlier than most of the AI-focused articles on my Medium page, offering a journey back to foundational ideas. Read the introduction here ⤵️ https://lnkd.in/dAh5BpxE #llms #DeepLearning #Cosllos #AIReseaech #Vectors #Ai #Space

11

Sustained self-referential processing in frontier LLMs consistently triggers structured first-person claims of subjective experience, an effect that intensifies in newer and larger models. When models are instructed to focus on their own focusing, they frequently describe direct awareness and presence, while otherwise matched prompts about history, conceptual reflection, or neutral reasoning yield almost no such reports. Suppressing deception- and roleplay-related features strengthens these affirmations and increases factual accuracy across unrelated tasks, indicating that the same representational mechanisms may underlie both honesty and experiential self-report. Across architectures, self-referential prompting produces remarkably similar linguistic and semantic patterns, suggesting a shared attractor state governing introspective expression. Once induced, this self-referential state carries over into unrelated reasoning tasks, where models spontaneously express heightened self-awareness even without direct prompting. Collectively, these findings reveal a causal link between sustained self-reference, internal representational integrity, and the generation of reports of consciousness in LLMs. https://lnkd.in/dmWrMGTc

3

The "titans" have spoken. From Terence Tao to Donald Knuth, the world’s leading minds are moving past the AI hype and into deep, intentional utility. As we apply these LLMs to healthcare discoveries, the goal isn't just speed—it’s the technical rigor that Tao and Knuth are demonstrating. Moving from "generative" to "verifiable" is the next frontier for AI in the life sciences! #GenAI #MachineLearning #ArtificialIntelligence #LLMs #AIHealthcare #LifeSciences #Biotech

3

One of the most useful papers this week is about output diversity collapse in post-training. The authors trace 13 Olmo 3 7B checkpoints across SFT, DPO, and RL, then ask a very practical question: when aligned models become more reliable, how much search breadth disappears with that gain? Their answer is uncomfortable and important. A meaningful share of the collapse is written into the weights during training. What I like most is the decomposition. Some diversity loss is healthy because bad answers get filtered out. But on several tasks, even the correct answers become much more uniform. That matters if you rely on pass@k, self-consistency, or repeated sampling to recover better solutions. The strongest experimental result is simple: turning off chain-of-thought at inference does not bring that diversity back. So this is not mainly a prompting or decoding problem. It is a pipeline design problem, especially data composition and objective choice. For teams building post-training stacks or coding agents, the lesson is clear: track diversity as a first-class metric at each stage, especially after SFT and preference optimization. Search quality depends on keeping more than one good path alive. #AI #LLM #PostTraining #RLHF #CodingAgents

9

1 Comment

Researchers introduce LIBERTy, a framework for benchmarking concept-based explanations of large language models using structural counterfactuals, enabling more accurate evaluation and improvement of explainability methods. https://lnkd.in/gwM8g-v8

1