Today's AI isn't just short on brainpower; it's missing specific machinery, like knowing what it doesn't know about you.
Which specific capabilities must AI develop beyond current language model abilities?
This explores which concrete abilities today's language models are missing, meaning specific gaps rather than 'more scale'. The corpus points to a cluster of missing mechanisms, not one missing skill.
This explores which concrete abilities today's language models are missing, meaning specific gaps rather than 'more scale'. The corpus points to a cluster of missing mechanisms, not one missing skill. Several things that look like weak intelligence turn out to be missing machinery for storing, checking, or acting on information.
The first gap is knowing what they don't know. Assistants have no representation of what remains unknown about the user, which helps explain sycophancy and hallucination. Adding a schema of labeled unknowns to prompts cut harmful advice and sycophancy by 50–75% and roughly halved hallucination Do language models know what they don't know about users?. A related gap is using what's in front of them. When training associations are strong, models produce outputs that contradict their own context. Better prompting doesn't fix this, and the fix appears to require intervening in the model's internal representations Why do language models ignore information in their context?. Memory is a third, architectural gap. Titans separates short-term attention from a long-term neural memory that stores the surprising tokens, and it scales past two million tokens without quadratic cost Can neural memory modules scale language models beyond attention limits?.
The next gap is actually learning new things. Post-training mostly selects reasoning that base models already hold rather than creating it. Five different elicitation methods all surface the same latent ability Do base models already contain hidden reasoning ability?. Prompting has the same limit: it can reorganize existing knowledge but can't supply knowledge the model never saw Can prompt optimization teach models knowledge they lack?. One route around this is Transformer2, which tunes only singular values in the weights so that expert skills can be mixed at inference time, allowing continual specialization with fewer parameters than LoRA Can models dynamically activate expert skills at inference time?. Self-improvement has its own ceiling, because it is formally bounded by the gap between generating an answer and verifying it What limits autonomous capability in large language models?. A model can't bootstrap past what it can check.
Another gap is acting in the world. Tool use isn't a convenience. There is a formal proof that tool-integrated reasoning enables strategies that are impossible, or absurdly verbose, in text alone Do tools actually expand what language models can reason about?. Turning that into autonomy takes more than a bigger model. Agent performance depends on training environments that scale in complexity, diversity, and real-world fidelity all at once, and a deficit in any one collapses generalization What blocks scaling from language models to autonomous agents?. Where real environments are scarce, a language world model trained to predict the next state can stand in for them. On three benchmarks it outperformed training in real environments Can language models learn to simulate agent environments?.
The last gap is creativity. Cognitive science distinguishes combinational, exploratory, and transformational creative reasoning. Existing LLM reasoning methods address only conventional problem-solving, which may explain why ideation collapses into similar answers Can LLMs reason creatively beyond conventional problem-solving?.
These findings come from separate papers, not one roadmap, and the corpus doesn't rank them. Taken together, the missing pieces are a sense of one's own ignorance, faithful use of context, durable memory, a way to acquire new knowledge, tools and worlds to act in, and structured creativity.
Sources 11 notes
Research shows assistants suffer from sycophancy and hallucination because they have no representation of what remains unknown about users. Adding a schema of labeled unknowns to prompts reduced harmful advice and sycophancy by 50–75% and cut hallucination rates by roughly half.
Research demonstrates that LMs generate outputs inconsistent with their context because parametric knowledge from training dominates over in-context information. Textual prompting alone cannot override strong priors; causal intervention in representations is required.
Titans architecture separates attention (short-term, quadratic) from neural memory (long-term, compressed), prioritizing surprising tokens for storage. The model outperforms standard Transformers and linear RNNs across tasks while scaling to 2M+ token contexts without quadratic penalties.
Five independent mechanisms—RL steering, critique fine-tuning, decoding changes, SAE feature steering, and RLVR—all elicit reasoning already present in base model activations. Post-training selects rather than creates reasoning; the bottleneck is elicitation, not capability acquisition.
Prompting works entirely within a model's pre-existing training distribution and cannot supply domain knowledge absent from training data. This creates a hard ceiling: no prompt strategy can compensate for missing foundational knowledge, only reorganize what already exists.
Show all 11 sources
Transformer2 demonstrates that tuning only singular values within weight matrices produces composable expert vectors that dynamically mix at inference without interference, outperforming LoRA with fewer parameters and enabling continual specialization.
Multi-agent deliberation produces specific failure modes (Degeneration-of-Thought, Silent Agreement), alignment at scale includes problematic self-valuation, and self-improvement is formally bounded by the generation-verification gap. Measurement error and conditional compliance hide the true capability ceiling.
Formal proof shows tool-integrated reasoning enables strategies impossible or prohibitively verbose in text alone, expanding both empirical and feasible support. The advantage spans abstract reasoning, not just arithmetic, and Advantage Shaping Policy Optimization stabilizes training without reward distortion.
Nex-N1 shows that autonomous agent performance depends on environment scaling along complexity, diversity, and real-world fidelity — not model size. Deficits in any single dimension collapse generalization, but scaling all three together enables frontier performance.
Qwen-AgentWorld demonstrates that native language world models trained via next-state prediction on 10M+ trajectories outperform real-environment training on three benchmarks and transfer across seven domains, positioning next-state prediction as a foundation objective for agents.
Research identifies combinational, exploratory, and transformational reasoning as distinct creative modes grounded in cognitive science. Existing LLM reasoning methods address only conventional problem-solving, leaving creative paradigms unaddressed and potentially explaining diversity collapse in ideation.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Eliciting Reasoning in Language Models with Cognitive Tools
- The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity
- Learning To Retrieve Prompts for In-Context Learning
- Graph Engineering in the Era of LLM Agents: From Individual Intelligence to System Intelligence
- Qwen-AgentWorld: Language World Models for General Agents
- Titans: Learning to Memorize at Test Time
- Universe of Thoughts: Enabling Creative Reasoning with Large Language Models
- Understanding Tool-Integrated Reasoning