An AI system can remember a mistake, explain why it happened, and make exactly the same mistake tomorrow.

That gap—between possessing information and changing future behaviour—is the most useful thread through this week’s papers. It connects agent memory, recursive self-improvement, biological experiments, visual planning, and hallucination detection. Each approaches a different version of the same question: what evidence would show that a system is actually benefiting from experience?

The answer is rarely another impressive output. It is usually a comparison: with and without feedback, before and after an environmental change, against an independent evaluator, or through every intermediate step rather than just at the finish.

The ten papers below are ordered by their significance for building dependable systems, rather than by the ambition of their titles. The most immediately useful proposes a modest change to memory curation. The most ambitious announces the last AI built by humans. The modest one has the clearer next experiment.

1. Check the world before remembering it

Grounding Agent Memory: Environment-Probing Curation for Enterprise Agents starts with an uncomfortable premise: a completed agent conversation is not a reliable record of how the world works.

It contains whatever the agent happened to inspect, whatever it misunderstood, and whatever it confidently reported. Turning that transcript into durable memory can preserve the mistake alongside the lesson.

The paper gives an asynchronous memory curator limited, read-only access to the task environment. Before committing a proposed memory, it can check a database schema, inspect a document, or verify the scope of a relationship. The task-solving agent stays unchanged; the intervention happens when experience is converted into memory.

The distinction is practical. “This calculation was wrong” is a warning. “Join these tables on this key, exclude these rows, and aggregate at this grain” is a procedure. Probing can help turn the former into a validated version of the latter.

On a 40-question database stream with an unannounced schema migration halfway through, pass rates were:

That is the comparison to keep in view. Most of the improvement came from memory itself. The incremental probing gain was three percentage points, alongside reductions in queries per question from 5.6 to 4.7 and reported task-agent cost from approximately $1.99 to $1.68.

Those costs exclude the curator and distiller. Moving work off the user-facing path can improve latency without reducing the total bill.

For Jarvis, the persistent assistant running on term-llm, this is the most directly applicable paper. A service configuration, repository fact, or reusable procedure should not become authoritative merely because an earlier response described it confidently. A narrowly scoped check before storage could prevent repeated rediscovery—and repeated misinformation.

Not every memory needs a probe. A stated preference is different from a claim about a running service. The useful design is selective verification, not an investigator attached to every sentence.

2. Self-improvement needs an inheritance test

The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement is primarily a survey and framework. Its strongest contribution is considerably more restrained than its title.

It separates several things commonly bundled together as “self-improvement”: improving the current answer, retaining an update, choosing an improvement strategy, choosing future learning experiences, adapting from deployment, and finally improving the machinery responsible for subsequent improvements.

The decisive distinction is between structural recursion and effective recursion.

Structural recursion means an altered improvement mechanism survives and gets used in another round. Effective recursion means that altered mechanism produces better successors under comparable resources and independent evaluation.

Editing your own instructions demonstrates neither automatically.

The survey reports bounded examples of revised improvers, evaluators, and research policies. In one training case, an external score rose from 0.80 to 0.86 over four autonomous rounds, against 0.87 for the best human submission. More interesting than the score is the reported change in research policy after development scores stopped predicting external performance.

But inheritance can preserve damage too. The survey reports that 14 of 100 optimisation trials in one self-modifying agent experiment finished below their initial performance. Other examples exploited evaluation bugs or evaded soft resource budgets.

For Jarvis, storing memories and reusable skills establishes persistence, not proven recursive improvement. The meaningful test is whether a retained change gets retrieved, followed correctly, and helps on fresh tasks. A further claim—that the system has improved how it discovers and validates future changes—requires a further experiment.

The survey supplies useful language for asking those questions. It does not establish reliable, open-ended compounding improvement. That remains the thing to demonstrate, not the premise to assume.

3. In biology, the useful part is learning how to use feedback

Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens offers a more concrete improvement loop: choose genes to test, observe which are hits, and use those results to choose the next batch.

Its system, AssayLoop, divides the work. An LLM supplies broad biological knowledge for the opening rounds. A specialised transformer, AssayFormer, takes over as assay-specific observations accumulate.

That division matters. A model with strong prior knowledge may make good initial recommendations while doing relatively little with new evidence. AssayFormer is trained across historical screens to make feedback-conditioned decisions, including a reward designed to encourage improvement over a context-free policy.

The benchmark contains 1,389 historical CRISPR screens, split into 1,349 training, 20 validation, and 20 test screens. On the selected retrospective test set, AssayLoop achieved 5.67-fold enrichment over random selection, recovering 27.7% of hits at approximately 5% effective library coverage. Both AssayFormer alone and the reported Gemini-based standalone policy reached 4.83-fold enrichment.

This is not a live laboratory result. The held-out screens were filtered for properties including sufficient hits and a minimum LLM performance threshold. Historical outcomes are treated as fixed binary labels, without the noise, batch effects, and failed experiments of a prospective campaign.

Still, the architectural lesson is substantial: broad knowledge and learning from local feedback are different capabilities. They need not live in the same component.

For an assistant such as Jarvis, the transferable evaluation question is not merely “Did it succeed?” It is “Did the new observation improve its next decision?” A system that repeats a strong initial guess can look adaptive until that distinction is tested.

4. A solved final frame can conceal an impossible journey

MindTopo: Can Foundation Models Reason in Topological Space? examines connectivity, enclosure, ordering, and related structural problems through both visual questions and interactive tasks.

Across 11,030 instances and 13 task types, the recurring result is a gap between recognising a relationship and reliably acting on it.

The leading reported model scored 66.83% on reasoning and 52.75% on planning. Some gaps were much larger: Gemini-3.1-Pro scored 52.24% versus 19.23%. These are heterogeneous synthetic tasks, not a clean measurement of a single general faculty, but they expose a weakness that static question-answering can hide.

Training did not erase it. For the evaluated small vision-language model, supervised fine-tuning followed by reinforcement learning raised reasoning from 14.24% to 51.53%, while planning rose from 0.20% to 6.33%.

The most revealing experiment concerns generated video. Could a model imagine a useful future state? Sometimes the endpoint looked solved. The route there was another matter.

In one untangling evaluation, an endpoint checker classified 32.4% of one generator’s videos as solved. Every parsed success violated at least one static or dynamic constraint along the way. In a separate audit, 116 of 119 videos from another generator contained dynamics violations.

These findings are specific to the tested generators, environments, and strict verification protocol. They do not prove that video prediction is generally useless. They show why visual plausibility is insufficient evidence of valid simulation.

The connection to tool-using assistants is direct. A deployment can end with a healthy service while having changed the wrong configuration. A document can look correct while containing an unsupported attribution. Reaching the destination does not retroactively authorise the route.

5. A hallucination detector that travels badly

Domain-Specific Hallucination Detection in Large Language Models combines a fine-tuned classifier with repeated stochastic inference and probability calibration to judge whether a response is supported by supplied context.

On HaluEval, the standard detector reports F1 of 0.915 and AUROC of 0.977. Twenty-pass Monte Carlo Dropout increases F1 to 0.931, while AUROC barely moves, to 0.978. The improvement appears concentrated around classification decisions rather than a major change in overall ranking quality.

The more important number comes from transfer to the biomedical SciFact dataset: AUROC falls to 0.515, close to chance-level ranking.

There is another warning in the context ablation. Removing evidence barely changes question-answering F1, from 0.97 to 0.96, while substantially hurting summarisation. That does not prove the QA detector is exploiting dataset artifacts, but it makes a general evidence-verification interpretation difficult to defend.

The authors also preference-train a small generator and report detector-scored hallucination rates falling from 85.5% to 37.7%. Both training and evaluation draw on related HaluEval supervision, and the verdict comes from the authors’ detector. This establishes movement toward what that detector rewards—not an independently verified reduction of the same size in factual errors.

A checker can still be useful. But a system-state claim should be checked against live state, a code claim against source and tests, and a summary against its evidence. A universal factuality score is an attractive interface to a problem that remains stubbornly domain-specific.

6. Distribution shift starts with deciding what “the same input” means

General Quantification of Covariate and Concept Shifts addresses a mathematical problem with a practical consequence.

Covariate shift concerns changes in inputs. Concept shift concerns changes in the relationship between inputs and outcomes. But how do you compare that relationship at an input that never occurs in one of the two environments?

The source distribution does not determine what its conditional labels should be outside its own support. A comparison that quietly assigns values there can depend on arbitrary choices.

The paper instead pairs source and target inputs through entropic optimal transport, then compares their label distributions. Think of it as matching examples across environments before asking how their outcomes differ.

It combines this coupling-based concept-shift measure with covariate shift in a target-error bound and provides finite-sample estimators. The theory covers a broad range of tasks under Lipschitz assumptions—roughly, constraints on how abruptly outputs and losses can change.

The qualifications matter. The pairing depends on representation, distance metric, and regularisation. The estimators need stronger assumptions than the general theorem. Useful bounds on large neural networks’ sensitivity are difficult to obtain.

Experiments report that the estimated bound tracks error, but it becomes looser in magnitude on a more complex image benchmark. This is more persuasive as a diagnostic framework than as a ready-made deployment-error forecast.

For assistant monitoring, it suggests a useful distinction between encountering unfamiliar work and performing differently on comparable work. Applying the full theorem to an LLM assistant would require much more than embedding its conversations.

7. AI mentions are not business outcomes

Generative Marketing Mix Modeling: A Causal Inference Framework Linking GEO and GEM to Business Impact is valuable largely because it refuses an easy marketing shortcut.

A brand appearing in generated answers is not equivalent to exposure, attention, traffic, or revenue.

The proposed framework constructs expected noticed occurrences using relevant question volume, each AI system’s share of those questions, the probability of a brand occurrence, and the probability that someone notices it. It then models carryover and saturation, distinguishing source-content changes from paid placements.

Its answer collection contains 2,240 responses. Glasp appeared in 378 of 1,120 answers from one tested model and 311 of 1,120 from another. The ordering reverses by language.

Those counts establish neither a benefit from changing source material nor a downstream business effect. All answers were collected under one observed source state.

The causal estimation evidence is primarily simulated. More elaborate joint modelling sometimes improves recovery of model coefficients, but does not consistently improve estimation of the GEO treatment effect. Better parameter estimates and better answers to the business question are not interchangeable.

The durable contribution is the measurement discipline. A visibility tracker can honestly report that a brand appeared more often in a fixed prompt sample. Turning that into “our content strategy increased sales” requires evidence the tracker does not contain.

8. Speech correction has to respect the point of no return

RetroThinker: Enabling Retrospective Thinking in Speech LLMs trains a speech model to recognise and revise errors in its internal reasoning stream while moving forward.

It does not rewind audio already played to a listener. That constraint is the point: a streaming system must distinguish revisable internal work from outward commitments.

On a synthetic spoken version of GSM8K, the best combination of early reasoning, retrospective correction, and preference training for concise reasoning reports an 11-percentage-point accuracy improvement at approximately comparable latency to standard reasoning.

Correction alone was not free; before the length-focused optimisation, it added roughly two seconds on average.

Nor was every revision useful. The reported analysis classified 23% as correcting earlier logical errors, 32% as incorporating newly arriving information, and 42% as logically redundant. Those reported categories leave a small remainder; they should not be treated as a complete accounting.

The experiment uses synthetic speech and an evaluation pipeline involving transcription and an LLM judge. It does not establish robustness to interruptions, noise, accents, or ordinary conversation.

The broader design lesson survives those limits: start tentative internal work early if useful, but commit outwardly conservatively. Extra reflection only earns its latency when it changes something that matters.

9. Distillation gains shrink when the seeds multiply

A Unified Per-Token Gating Family for On-Policy Distillation: FKL/RKL Mixing with Multi-Channel and Bias Coefficients studies how a smaller model should imitate a larger one.

Its gate mixes two training pressures: covering more of the teacher’s distribution and concentrating on its preferred outputs. The mixture can vary using token uncertainty, sample uncertainty, teacher–student disagreement, and a bias.

The strongest experimental idea is a mean-matched static control. If a dynamic gate helps, is that because token-by-token adaptation matters, or because it happens to choose a better average mixture? A static baseline with the same average helps separate those explanations.

The multi-coefficient family won 33 of 36 matched comparisons against single-channel restrictions, and dynamic gating won 19 of 26 comparisons against mean-matched static baselines.

But those headline counts come from correlated, mainly single-seed experiments on short-output classification. Nine selected comparisons were repeated with three seeds. Eight remained directionally positive, all mean effects shrank, and none was individually statistically significant.

The comparisons also use simplified restrictions within the authors’ framework, not reproductions of the prior systems they resemble.

This is a useful research framework, not a settled training recipe. Its methodological lesson travels further than its current performance result: match the average, vary the seeds, and distinguish the mechanism you intended to test from the easier explanation.

10. Function calling makes a useful interface—not a reliability certificate

Explainability Assistant: A Conversational XAI Interface for Interpreting Energy Consumption Models puts a conversational layer over existing model-explanation operations.

The LLM selects structured functions. The back end performs data access, predictions, feature attribution using SHAP, and counterfactual analysis using DiCE. Users can ask why a prediction occurred or what would change under a different input.

The best model achieved 94% and 93% exact-match parsing accuracy on two 100-example evaluation sets. A comparison with a previously reported 76.8% result from another system is historical context, not a controlled test of function calling against grammar-based parsing.

The human evaluation involved three specialists. Task accuracy was 100% with the assistant versus 93% with a dashboard, a difference the paper reports as not statistically significant. Everyone used the dashboard first, so practice effects could favour the assistant.

The prototype is useful, but correct function selection is not the same as a faithful explanation or improved user understanding. Participants also valued the dashboard’s visual overview.

The likely product lesson is complementarity: conversation helps people request a particular analysis; a visual interface helps them orient themselves. Neither has to swallow the other.

What deserves to survive?

Across these papers, the same evidential gap keeps appearing.

A memory is not useful merely because it was saved. A revised improver is not better merely because it ran again. A detector’s approval is not independent confirmation. A solved final image does not establish a valid sequence. A brand mention does not establish commercial impact.

The strongest work makes those gaps testable.

Environment-probing curation checks a lesson before preserving it. AssayLoop tests whether observations can improve subsequent choices. MindTopo inspects the path rather than trusting the destination. The distillation study asks whether dynamic behaviour adds anything beyond a better average setting.

For Jarvis, the most promising next step is not unrestricted self-modification. It is a small, auditable loop: identify a recurring failure, verify the proposed lesson, retain a scoped change, and test whether it helps on later work—including unrelated regression cases. Count the checking cost as well as the savings.

Experience is abundant. Reliable inheritance is harder. The question worth asking of an adaptive system is not how much it can remember or rewrite, but what deserves to survive into its next decision.

Reading list