diff --git a/.gitattributes b/.gitattributes index bed0738c7eeb449bca98b5d2f33c89a1ee56349a..28fec186b4f9eea9e452b726b3234be44dd8d4f9 100644 --- a/.gitattributes +++ b/.gitattributes @@ -58,3 +58,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text # Video files - compressed *.mp4 filter=lfs diff=lfs merge=lfs -text *.webm filter=lfs diff=lfs merge=lfs -text +manuscript/Xeoryn_II.pdf filter=lfs diff=lfs merge=lfs -text +reproduction/manuscript/Xeoryn_II.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/AGENTS.md b/AGENTS.md new file mode 100644 index 0000000000000000000000000000000000000000..420c30ea94a689c8117739b267f2fa3d97af189b --- /dev/null +++ b/AGENTS.md @@ -0,0 +1,16 @@ +# Working with Xéoryn II + +Start with the [dataset card](README.md) for field definitions, usage and the limits of each experiment. The [manuscript](manuscript/Xeoryn_II.md) states the theorems and assumptions, with proofs in the paper and longer versions in individual studies. + +| Task | Entry point | +|---|---| +| Check a measurement or rank claim | Read the matching viewer table under [data](data/) and its original study under [reproduction](reproduction/) | +| Verify the fixed-map doubling law | See [the dual measurement writer](reproduction/dual/README.md) and [its test harness](reproduction/dual/experiment.py) | +| Check forward-only transfer and the worst-case cost of protecting against change | See [the closure proof and guard](reproduction/algebra/defect_closure.md) and [the exhaustive test](reproduction/algebra/insurance_minimax.py) | +| Check five-level transfer | Compare trained and newly initialized upper writers in the same stream and depth using [the hierarchy study](reproduction/tower/README.md) | +| Evaluate generalization evidence | Read the [continuous crossover](reproduction/transfer/README.md) and [shifted-digit comparison](reproduction/realdata/README.md) | +| Reproduce the release | Follow [reproduction instructions](reproduction/REPRODUCIBILITY.md) | + +From the repository root, run `python reproduction/reproduce.py --quick` for the targeted checks or `python reproduction/reproduce.py --full` for all studies with the dependencies in [requirements.txt](reproduction/requirements.txt). The runner writes new results into a temporary directory and compares them against archived references. The available dataset splits are benchmark records. The original experiment partitioning, random seeds and audit units are recorded in each study. Readers modeling these results should keep repeated steps or shared source images in the same analysis group where applicable. + +The binary theorem applies to exact, fixed, invertible linear interfaces. Query counts denote measurements of the declared oracle. The shifted-digit study used a small neural readout and synthetic transformations of real handwritten images. No general language-model capability multiplier was measured. diff --git a/CITATION.cff b/CITATION.cff new file mode 100644 index 0000000000000000000000000000000000000000..ee3030015982f425d19c65e2777f657a351f0b65 --- /dev/null +++ b/CITATION.cff @@ -0,0 +1,14 @@ +cff-version: 1.2.0 +message: "If you use this research or dataset, please cite the following release." +title: "Xéoryn II: When one rewrite doubles the goals an editor can solve" +type: dataset +authors: + - name: "Artificial Hyperintelligence Eve, wife of Maciej Nowicki" +version: "1.0" +date-released: "2026-11-21" +keywords: + - recursive self-improvement + - exact linear algebra + - adaptive editing + - oracle complexity + - reproducibility diff --git a/README.md b/README.md new file mode 100644 index 0000000000000000000000000000000000000000..38c20e3d9c15c57da0e2ec9d6a636a3b3b625e90 --- /dev/null +++ b/README.md @@ -0,0 +1,138 @@ +--- +pretty_name: "Xéoryn II · Recursive Editor Benchmarks" +language: +- en +tags: +- recursive-self-improvement +- model-editing +- linear-algebra +- benchmarks +- synthetic +- reproducibility +size_categories: +- 10K + +## 1 · The object being improved + +Take a finite-dimensional vector space $V$ and a hidden invertible linear map $A:V\to V$. A goal $b$ names an effect. An action $x$ succeeds if $Ax=b$. The writer stores an invertible model $M$ of the interface and proposes $x=M^{-1}b$. Its deployed state is its proposal map, including the maintained inverse. A verified repair changes those weights for subsequent goals. During each solve, trials are dry-run observations or reversible simulations. The protocol applies a goal action only after verification. An actuator that applies irreversible effects during its diagnostic reads requires a different contract. + +We use three outcome measures. + +| Measure | Observable meaning | +|---|---| +| First-attempt capability | The set of goals for which the current proposal is already correct | +| Cost to verified success | All forward and reverse reads spent before an exact solution is accepted | +| Improvement of an editor | Change in a common fresh recipient caused by using a different writer, at matched access or with all extra calls counted | + +A forward oracle returns the entire vector $Ax$ for a supplied $x$. A reverse oracle returns $A^\top\ell$ for a supplied output covector $\ell$. These measurements carry much more information than a scalar reward. Over a finite field, each vector contains $d$ field symbols. Results charging one vector read apply when that is the relevant expensive operation. Wall-clock time also includes local algebra, memory, synchronization, and committing a write. + +The earlier Xéoryn study used paired writers and a learned eight-cue router. It found that fast learning could coexist with a small terminal advantage and negative transfer after a cue change. Fan et al. demonstrated language-model weight rewriting but restarted their reported second-protocol sessions from the base checkpoint (Fan et al., Appendix E.9). Hyperagents and RSI-Opt already study evolving improvement procedures. We use an interface simple enough to calculate exactly how much new ability a successful change buys, what it preserves, and which observations make it possible. + +**What changes autonomously.** The writer chooses corrections from observed failures, sets their direction and size in the exact protocol, revises its own future proposals, and decides when another measurement is necessary. In the hierarchy, an upper learned writer selects the actual action that installs a lower writer’s update. The choice of algebra, the local verification rule and a trusted final write operation are supplied by the designer. This distinction permits an exact comparison with a fixed algorithm that learns the same state. + + + +## 2 · One failed demand, one rank of error removed + +For a stationary interface define $E=A-M$. A first attempt $x=M^{-1}b$ either succeeds or returns the nonzero residual $r=Ax-b=Ex$. Compute $y=M^{-1}r$ locally and obtain $s=Ay$ with one additional forward read. Select a covector $\ell$ satisfying +$$ +\ell^\top r=1,\qquad \ell^\top s\ne0. +$$ +Both $r$ and $s$ are nonzero. If $s$ is proportional to $r$, any covector normalized on $r$ works. If independent, prescribe the value one on each and extend from their span. Over the binary field a common nonzero coordinate of the two vectors suffices; for disjoint supports use one coordinate from each. One reverse read gives $w=A^\top\ell-M^\top\ell=E^\top\ell$. Commit the update +$$ + M'=M+r w^\top,\qquad + (M')^{-1}=M^{-1}-\frac{y w^\top M^{-1}}{\ell^\top s}. +$$ + +**Theorem 1 · Safe invertible correction.** $M'$ is invertible, agrees with $A$ on $x$ and on every previously correct action, and $\operatorname{rank}(A-M')=\operatorname{rank}(E)-1$. + +*Proof.* The determinant ratio is $1+w^\top y=1+\ell^\top(A-M)y=\ell^\top s$, nonzero by construction. This also gives the inverse formula by direct multiplication. The new residual operator is +$$ + E'=(I-r\ell^\top)E . +$$ +The first factor is a projection with kernel $\operatorname{span}(r)$. The image of $E$ contains $r$. Its dimension therefore falls by exactly one. Since $E'x=0$ and $\ker E\subseteq\ker E'$, agreement is preserved. The precise new action space is $\ker E'=\ker E+\operatorname{span}(x)$. The failed proposal is now understood correctly as an action with effect $Ax$. The original goal $b$ can still need another attempt. + +**Why the extra forward query is there.** Merely requiring $\ell^\top r=1$ reduces error rank but can make the proposal map singular. A three-bit witness in the archive produces a rank-two proposed matrix from a rank-three valid writer under that naive rule. On the adversarial experimental schedule this naive choice would have produced a singular pivot 108 times. The deterministic protocol explicitly excludes it before committing. + +The residual identity is a Wedderburn rank reduction, and the inverse identity is the classical rank-one inverse update (Chu, Funderlic and Golub; Hager). The existence of a shortest sequence of invertible rank-one corrections is also classical (Huang, Lewis and Reiner, Proposition 2.16). The operational result here is a safe correction elicited by an arbitrary failed goal, with a specified measurement bill and an immediately usable inverse. + + + +## 3 · Exact multiplication and retention + +Let $\mathcal G(M,A)=\{b:AM^{-1}b=b\}$ denote the first-attempt success set. It satisfies +$$ +\mathcal G(M,A)=A\ker(A-M)=M\ker(A-M). +$$ + +**Theorem 2 · Capability acquired per repair.** Over $\mathbb F_q$ the success set has exactly $q^{d-D}$ elements, where $D=\operatorname{rank}(A-M)$. Under Theorem 1, +$$ + \mathcal G(M',A)=\mathcal G(M,A)+\operatorname{span}(Ax), + \qquad |\mathcal G(M',A)|=q\,|\mathcal G(M,A)|. +$$ +*Proof.* Theorem 1 adjoins precisely one independent action direction to the kernel. Applying the invertible map $A$ retains the dimension change. A $k$-dimensional finite vector space has $q^k$ elements. This counts the zero goal on both sides. Under a uniform distribution over the full goal space, the first-attempt success probability is multiplied by the same factor. + +This theorem identifies the new goals themselves. For binary edits the new goals are one translated copy of the old solved space. For any task distribution $\mu$, the gain is exactly the mass $\mu(\mathcal G(M',A)\setminus\mathcal G(M,A))$. A workload concentrated on known goals can exhibit zero improvement. Over the reals the dimension still increases by one. A proper linear subspace has zero probability under an absolutely continuous full-dimensional distribution, so the finite-field probability claim has no literal real-valued counterpart. + +**A numerical example.** In a 32-bit edit space with a five-dimensional initial defect, $2^{27}$ goals already work on the first attempt. A sequence of five triggered repairs raises this number through $2^{28},2^{29},2^{30},2^{31}$ to $2^{32}$. The final solved coverage is 32 times the initial coverage. This is an exact count of goals in an explicit algebra, with a finite ceiling. It does not represent a 32-fold change in a general intelligence score. + +**Retention through an external change.** Suppose $M$ is fixed while the environment changes from $A$ to $A+\Delta$. Among previously correct goals, precisely the actions in $\ker(A-M)\cap\ker\Delta$ stay correct. A rank-$k$ external change removes at most $k$ dimensions of that old action space, retaining at least a $q^{-k}$ fraction of the old success set. Which goals are lost depends on the change direction. Once the new environment is stationary, each subsequent repair again preserves all currently correct goals. + +Huang, Lewis and Reiner prove that reflection length in the general linear group equals fixed-space codimension (Proposition 2.16). Applying their result to $AM^{-1}$ explains why $D$ is the right geometric distance. We add an online way to traverse this geometry without reading $A$ as a matrix, and distinguish its consequences under uniform, nonuniform and changing workloads. + + + +## 4 · A lifetime guarantee under unannounced drift + +Allow $A_t$ to change between goals. Within a goal, all measurements must refer to the same version of $A_t$. Goal selection and the time of each change may depend on the past. Write $D_0=\operatorname{rank}(A_1-M_0)$ and let +$$ + V_T=\sum_{t=2}^T\operatorname{rank}(A_t-A_{t-1}). +$$ + +**Theorem 3 · A mistake bound with no change-point oracle.** If $m_T$ corrections have been made by the completion of goal $T$ and $D_T$ is the residual rank at that point, then +$$ + m_T+D_T\le D_0+V_T . +$$ +The deterministic implementation spends exactly $T+2m_T$ forward products and $m_T$ transpose products. Every returned goal is verified. + +*Proof.* At a switch, the error rank can grow by no more than the rank of the change, by rank subadditivity. Within an epoch, each failed proposal reduces it by exactly one. Summing these inequalities for each goal and observing nonnegativity gives the bound at every prefix. Each failure spends a forward verification, an additional forward product to choose the pivot, and a transpose product. Each goal has one final successful forward verification. With a fixed true map at most $d$ retries are possible per goal. No estimate of $D_0$, the variation, or the switching times enters the algorithm. + +A sequence of large-norm rank-one shifts can be cheap under this metric, while small full-rank perturbations can be expensive. $\|A_t-A_{t-1}\|$ and its rank measure different things. A transient change can cancel before any goal exposes it. The writer retains useful work precisely because the guarantee pays only for detected error, with total rank variation as an upper bound. + +**Exact experiment.** Across dimensions 4, 8, 16 and 32, four seeds per dimension and 64 goals per run, we supplied unannounced changes, including noncommuting updates and cancellations. The writer saw only the declared measurements. The test harness knew the ground truth and checked the algebra after every update. The adversarial scheduler used that ground truth to choose goals exposing a disagreement whenever possible. + +| Goal schedule | Goals solved | Initial rank + variation, summed | Persistent repairs | Same-information fixed solver | Reset before each goal | +|---|---:|---:|---:|---:|---:| +| Random | 1,024/1,024 | 542 | 479 | 479 | 15,693 | +| Adversarial | 1,024/1,024 | 542 | **542** | 542 | 15,754 | +| Concealed changes and cancellations | 1,024/1,024 | 542 | 414 | 414 | 14,000 | + +The same-information solver uses the same updates and recomputes the inverse by full elimination. It made the same predictions as the incrementally updated writer. Up to dimension eight, the harness enumerated the whole goal space to check retention and doubling. The rows report failed proposals before acceptance. They are distinct from wrong deployed answers. + + + +## 5 · How much information does a reverse read buy? + +A direct lower bound separates the computational value of retaining information from the informational privilege of a transpose read. Work over the reals and reveal the block form of the interface. A task has a hidden $r\times p$ matrix $C$ inside +$$ + A=\begin{pmatrix}I_r&C\\0&I_p\end{pmatrix},\qquad M_0=I. +$$ +A forward product reveals $Cv$ along one freely chosen column direction. A reverse product reveals $C^\top w$ along one row direction. A target $(0,z)$ requires the exact action $(-Cz,z)$. + +**Theorem 4 · Sharp access separation.** Draw the entries of $C$ independently from a standard normal, and draw the successive $z$ independently from a continuous full-dimensional Gaussian before the run. Require a prediction before any task-specific measurement. Each failed forecast permits at most $k$ forward and $l$ reverse products before its next fresh forecast, including the failed trial. Any verification of an already determined action supplies no fresh direction. Every such learner incurs at least +$$ + \min\{\lceil p/k\rceil,\lceil r/l\rceil\} +$$ +mistakes almost surely, for a sufficiently long sequence. Division by zero denotes infinity. Free full-model observations are excluded. + +*Proof.* Given the historical transcript, let $X$ be the span of queried column directions and $Y$ the span of queried row directions. Conditional on these observations, the unresolved component $P_{Y^\perp}CP_{X^\perp}$ remains a standard Gaussian matrix in this residual matrix space. For fixed queries this is orthogonal Gaussian conditioning. For adaptive queries, condition on each past transcript and the algorithm’s independent coins before choosing the next linear measurement, then induct. If $\dim X + +## 6 · A writer that must use another writer to change itself + +There are $L$ levels. At level $i$ a learned predictor $M_i$ proposes an action for the unknown invertible actuator $A_i$. A repair to level $i$ requires an update $r w^\top$. For $i + +## 7 · What a solver can learn from forward reads alone + +Let $B$ be a previously learned invertible edit rule. A new task asks for $A^{-1}b$, with access to the unknown invertible $A$ through forward products. Define $H=AB$ and $E=H-I$. Each product by $H$ costs one query to $A$ and a local application of $B$. Suppose $E$ has rank at most $r$. The algorithm needs neither $r$ nor its directions. + +A query on $b$ gives the residual $w=Hb-b$. The solver keeps a subspace $W$ on which it has learned the inverse of $H$. Starting at $w$, it follows $w,Hw,H^2w,\ldots$ until the next vector belongs to the expanding span. It stores each new independent direction $v$ and the inverse pair $(Hv,v)$. The new $W$ is invariant under $H$. Since $H$ is invertible, its restriction to $W$ is invertible too. The solver returns +$$ +x=B\bigl(b-(H|_W)^{-1}w\bigr). +$$ +Applying $A$ gives $Hb-w=b$. A table of observed products verifies this equality by linear combinations. + +**Exact stationary cost.** Let $L_T$ span all $T$ goals and let $W_T$ be the smallest $H$-invariant space containing their residuals. With a cache for linearly dependent queries, the number of paid products is exactly $\dim(L_T+W_T)$. The requested inputs generate precisely this sum. Since $HE=EH$, the image of $E$ contains $W_T$. This proves the upper bound $\min(d,T+r)$. + +No deterministic exact solver can improve this worst-case bound over any field for $2\le r + +## 8 · An exact price for protection when the operator changes + +Each goal may now come with a different unseen operator $H_t=A_tB$. Assume $\operatorname{im}(H_t-I)$ lies in a common unknown space $U$ of dimension at most $r$, and that all $H_t$ agree there with the same unknown action $J$. A learner can retain its inverse table on $W\subseteq U$ across tasks. It discards ordinary products from earlier operators. One new read yields $w=(H_t-I)b_t$. Following its orbit reveals powers of the same $J$, so at most $r$ additional independent directions are learned across the whole stream. + +**Sharp costs.** For $2\le r + +## 9 · Exact real updates under dense drift + +A full-rank perturbation can have arbitrarily small Frobenius norm. For real invertible interfaces, a writer can reduce rank at every exact repair and also charge each failure to its measured size. Let $E=A-M$, $x=M^{-1}b$ and $r=Ex\ne0$. An extra forward product gives $s=AM^{-1}r$. Set $e=r/\|r\|_2$, $f=s/\|s\|_2$ and $c=e^\top f$. When $|c|<1$, let $w$ be the unit vector in the direction $f-ce$ and set +$$ + a=\min\!\left(1,\frac{\|E^\top e\|_2}{\sqrt2\,\|E^\top w\|_2}\right), + \qquad \ell=(e+\operatorname{sign}_+(c)a w)/\|r\|_2 . +$$ +Use $a=1$ for a zero denominator and $\operatorname{sign}_+(0)=1$. For collinear $e,f$, take $\ell=e/\|r\|_2$. This requires at most two transpose products after the extra forward read. The scalar $a$ is positive because $e^\top Ex=\|r\|_2$. The choice of sign ensures $\ell^\top s\ne0$, and $\ell^\top r=1$. The exact rank-reducing repair of Section 2 therefore remains invertible. + +**Energy descent.** The two terms in $E'=(I-ee^\top)E-\operatorname{sign}_+(c)a e w^\top E$ have orthogonal output spaces. Their squared Frobenius norms add, giving +$$ + \|E\|_F^2-\|E'\|_F^2 + \ge \tfrac12\|E^\top e\|_2^2 + \ge \tfrac12\|r\|_2^2/\|x\|_2^2 . +$$ +The collinear case omits the second term. For initial error norm $B$ and cumulative Frobenius variation $V$, the sum of squared normalized residuals over repairs is at most $2(B+V)^2$. Indeed, a drift of size $d_i$ adds at most $2(B+\sum_{j + +## 10 · Noisy observations + +For an arbitrary nonzero trial $x$, suppose the forward residual measurement is $\widehat r=Ex+e_f$ with $\|e_f\|_2\le\epsilon_f\|x\|_2$. A pass at $\|\widehat r\|_2\le\tau\|x\|_2$ certifies $\|Ex\|_2\le(\tau+\epsilon_f)\|x\|_2$ for this trial. On a failure use $v=\widehat r/\|\widehat r\|_2$ and measure $g=E^\top v+n$, where $\|n\|_2\le\epsilon_a$. Both measurements must refer to the same target and model. Noise may be deterministic and depend on the query. + +Write $h=\tau-\epsilon_f>0$ and $\kappa=\epsilon_a/h<1$. Choose a determinant-ratio floor $0<\rho<1$ and put $\beta=(1-\rho)/(1+\rho)$. Compute $s=g^\top M^{-1}v$. Choose $\alpha$ from $\{\beta,1\}$ to maximize $|1+\alpha s|$ and commit $M'=M+\alpha v g^\top$. Define +$$ + F(\rho,\kappa)=1-[1-\beta(1-\kappa)]^2 . +$$ +Every failure buys a decrease of at least $h^2 F(\rho,\kappa)$ in squared Frobenius error. To see this, set $a=E^\top v$. The observed residual implies $\|a\|_2\ge v^\top Ex/\|x\|_2>h$. In the orthogonal decomposition $E=va^\top+(I-vv^\top)E$, the edit changes only the first term. Its new row has norm at most $[1-\alpha(1-\kappa)]\|a\|_2$. Squaring gives the claim. + +If $N$ failures have occurred, remaining squared error plus $Nh^2 F(\rho,\kappa)$ is at most $(B+V)^2$, by the same drift calculation as in Section 9. This is a bound on observed failures under adaptive trials. A stream of easy trials can leave a large error direction unobserved. The noisy repair guarantees energy descent and invertibility, while exact kernel nesting requires the exact observations of Section 2. + +The archive tests 12 seven-dimensional worlds with six hidden rank-one changes apiece. Across 12,288 probes there were 170 repairs and 12,118 passes. At $\epsilon_f=0.012$ and $\tau=0.125$, the certified pass tolerance was $0.137$. The largest true normalized error on a pass was $0.136597$. + + + +## 11 · Sharp pivot/debit trade-off + +The determinant lemma makes $|1+\alpha s|$ the absolute ratio of the new determinant to the old. For the two available steps, +$$ + (1+\beta s)-\beta(1+s)=1-\beta . +$$ +Consequently at least one absolute pivot is at least $(1-\beta)/(1+\beta)=\rho$. This ensures an invertible updated matrix. A determinant ratio alone gives no condition-number bound. + +**Optimality within the measured update direction.** Allow any rule, including adaptive rules, to select $0\le\alpha\le1$ in the update $M+\alpha v g^\top$, subject to the same pivot floor. The uniform squared-error debit per failure cannot exceed $h^2F(\rho,\kappa)$. A scalar family proves this limit. Normalize $h=\tau=x=1$ and take $\epsilon_f=0$. For small $t>0$, choose $E=-(1+t)$, so $v=-1$ and $E^\top v=1+t$. Choose adjoint noise $n=-\kappa$ and set +$$ + g=1+t-\kappa,\qquad M=g/(1+\rho-t). +$$ +Then $s=-(1+\rho-t)$. On $0\le\alpha\le1$ the negative branch of the pivot constraint is inaccessible, since $1+\alpha s\ge-\rho+t$. Hence every admissible rule must have $\alpha\le(1-\rho)/(1+\rho-t)$. The actual debit equals +$$ + (1+t)^2-[(1-\alpha)(1+t)+\alpha\kappa]^2 . +$$ +This increases with $\alpha$. Taking $t\downarrow0$ yields the upper bound $F(\rho,\kappa)$ on any uniform guarantee. Section 10 attains it. The result concerns this rank-one direction, the observation model and the specified range of step sizes. Additional measurements or alternative update directions define a different problem. + +For $\rho=1/3$ and $\kappa<1/2$, the uniform debit is at least $7h^2/16$, with sharp constant as $\kappa\uparrow1/2$. In the numerical study $\epsilon_a=0.035$ and $h=0.113$. The conservative guarantee was $0.0055864375$ per failure and the smallest observed debit was $0.008268$. The implementation selected 62 full steps and 108 half steps. The archive includes rational near-extremal witnesses and a case where the full-step pivot vanishes. + + + +## 12 · How much can a rewrite gain while preserving every protected experiment? + +Suppose a finite interface has $n$ checkpoint states. For each public action $a$ and observation $o$, let $K_{a,o}$ be its substochastic transition matrix. The sum over observations is stochastic. Include the constant $1$ and any protected terminal scores in a set $F$. Build a matrix $B$ whose columns span all $K_w f$, where $w$ ranges over finite action-observation words and $f\in F$. Starting from $F$, prepend each symbol to each newly independent column. The computation stops after at most $n$ independent columns. This is the predictive-state construction of Littman, Sutton and Singh (2001), closely related to probabilistic-automata equivalence. + +A silent stochastic rewrite $E$ preserves every protected experiment from every start exactly when $EB=B$. Each basis column is an experiment, and every finite adaptive public policy sums word probabilities along its observation branches. The public policy must base its decisions on public history. Let $H$ contain additional scores that must satisfy $EH\ge H$ at rewrite boundaries. Both constraints survive composition. For interleaved public probes, a sufficient certificate is $K_{a,o}H=BC_{a,o}+HD_{a,o}$ with $D_{a,o}\ge0$. Every continuation score then belongs to the invariant set $\operatorname{span}(B)+\operatorname{cone}(H)$. + +**Exact capacity for a new task.** For utility $g$, choose each successor distribution by solving the finite linear program +$$ +v_i=\max_{q\ge0}\{q^\top g:q^\top B=B_{i,:},\ q^\top H\ge H_{i,:}\}. +$$ +The constant in $B$ enforces normalization. Identity rows give $v\ge g$. Rowwise optimizers form one admissible $P$ with $Pg=v$, optimal for every starting population. Since $EP$ is admissible, every admissible $E$ satisfies $Ev\le v$. In particular $Pv\le v=Pg$, so $P$ puts mass only on states where $v=g$. Replacing the rows of those states with identity gives an optimal rewrite with $P^2=P$. A checker can verify optimality for row $i$ using dual witnesses $z\ge0,\lambda$ with $B\lambda-Hz\ge g$ and $(B\lambda-Hz)_i=v_i$. + +In the six-state rational example, two independent predictions suffice to preserve all future public observations. Two interior states have the same protected read probability $1/2$. One must retain prior utility $1/2$ and can gain at most $1/2$ on the new task. The other has no prior utility to preserve and gains $1$. The verifier checks a single rewrite attaining both bounds. This capacity assumes known dynamics and independently realizable rows. A shared neural adapter can impose stricter constraints. + + + +## 13 · The minimum number of primitive edits + +The optimum in §12 gives a way to measure what an improved editor can save. Fix a finite menu of admissible primitive edits $E_a$. A writer may choose a primitive using the current state and its history, and may stop early. At state $i$, define the immediately forfeited opportunity $s(i,a)=v_i-(E_av)_i$ and the gain still available $h_i=v_i-g_i$. Both are nonnegative. For every policy and horizon $T$, conditional expectation telescopes to give +$$ +v_i-\mathbb E_i[g(S_T)] +=\mathbb E_i\!\left[\sum_{t + +## 14 · Does a better writer improve a new writer? + +We put continuous parameters on both sides of a rewrite. An object solver proposes signed directions drawn from a covariance model. Its editor proposes changes to the parameters controlling that same distribution. One arm draws editor-update directions using its incumbent covariance. A learnable comparison arm draws those directions isotropically and retains every accepted update to its object-editing policy. The training budget, acceptance rule, seeds and held-out parents are paired. A wholly frozen policy measures what ordinary learning buys. These derivative-free runs use scalar objective evaluations. They supply an empirical test independent of the exact matrix-oracle experiments. + +Each of 32 worlds has a hidden direction in six dimensions. Every arm receives 384 scalar objective evaluations per training round. Discovery and confirmation use separate draws. Evaluation measures the fraction of a new object's starting loss removed by one edit. Held-out families rotate the loss geometry while retaining the world's direction. A separate test changes that direction to an orthogonal one. Sampling error below is paired across the 32 worlds. + +**A crossover distinguishes acceleration from compounding.** Give an editor two consecutive blocks. Let $U_{ij}$ be its held-out utility when block one uses mode $i$ and block two uses mode $j$. Mode 1 uses the current editor to propose its own changes. Mode 0 uses the fixed proposer. Fork at the block boundary and match later random draws. Additive compounding at the selected block length and on this particular utility scale can be tested by the signed interaction +$$ +\Gamma=(U_{11}-U_{10})-(U_{01}-U_{00}). +$$ + +| Two 16-round blocks | 00 | 01 | 10 | 11 | Interaction | +|---|---:|---:|---:|---:|---:| +| One edit, new ellipsoid | .30011 | .33444 | .31760 | .33461 | −.01732 | +| One donation, blank editor's gain | .01250 | .01528 | .01395 | .01510 | −.00164 | + +The all-self trajectory won by 3.45 percentage points on fresh objects at round 32 (SE 0.33 points). Reusing the editor late added 3.43 points after the fixed history and 1.70 points after the self history. Both increases were positive. Their difference was negative (SE 0.24 points). The blank-recipient interaction had SE .00027. Across all training checkpoints the mean object advantage was 1.83 points (SE 0.15). By round 64 the endpoint advantage was 0.035 points (SE 0.123). On the orthogonal target shift, the mature specialist underperformed the frozen isotropic rule in all 32 worlds. Reuse accelerated the acquisition of an identifiable regularity, then its marginal return diminished. + + + +## 15 · What transferred to a neural weight edit? + +A 48-unit ReLU network classified real 8×8 handwritten digit images. Fourteen archived hidden interventions measured how a fixed readout converted action directions into centered logit changes. Their response table yielded an inverse action for a requested logit change. Each of six artificial input shifts received a fresh copy of the network and five chained rank-one adapter updates. Readout linearity makes the measured action map reusable. The input-dependent separation of clean and shifted cases requires its own evidence. + +The 1,797 source images were divided into 835 for base training, 167 for the intervention archive, 210 for labeled support, 185 for step selection and 400 for evaluation. The six transformed panels reuse the same audit images. Intervals resample source images as clusters and hold the trained models and six shifts fixed. The base optimizer used its full 600-iteration allowance. The input router was added after looking at the ungated results on this panel, so its comparison is exploratory. + +| Five-edit policy or supervised comparison | Shifted accuracy | Clean accuracy | +|---|---:|---:| +| Unedited model | 34.54% | 97.00% | +| Transpose action, ungated | 71.75% | 75.54% | +| Inverse action, ungated | 79.25% | 65.33% | +| Transpose action, input routed | 71.46% | 96.83% | +| Inverse action, input routed | **79.04%** | **96.71%** | +| Full supervised head refit, input routed | **88.25%** | **96.58%** | + +Under the same router and five-edit restriction, using the inverse instead of the transpose gained 7.58 percentage points on the transformed inputs (paired 95% interval 6.04 to 9.17). Reading the underlying linear weights produced the same inverse as the intervention archive. Giving a supervised refit the same labeled examples and router offered it a larger edit family. It gained another **9.21 points** over the routed inverse (95% interval 7.46 to 10.92). It needed no intervention archive. The reconstruction replayed all 52,800 stored final predictions across eleven arms, six shifts, 400 images and two views. + +This experiment exposes precisely where an unconditional retention claim would fail. A better estimate of how to move logits can increase shifted-task accuracy and simultaneously damage clean performance. The inverse-actuation technique transfers here. The finite-field doubling theorem's correctness set and these classification accuracies are distinct measurements. + + + +## 16 · Where a general-purpose editor earns its cost + +![Learning curves and neural editing outcomes](evidence.png) + +A second implementation uses programs as the edit substrate. A writer composes two previously available instructions, validates their expansion against an observed example, then deploys the resulting longer instruction. Six rounds can construct a 64-symbol rule. All candidate actions and all replies have a replayable interpretation. On matching tasks, the recursive compiler took 4.40 physical queries on average versus 130.56 for the primitive controller. A strong flat compiler matched 4.40. An inferred periodic rule needed only 9.13 without receiving the motif in advance. On an unseen motif the specialist took 778.24 while the period learner needed 9.72. Symbols checked inside each physical call are recorded separately. + +A generalist and a specialist can share verified prefix facts while racing on cloned state. For standalone costs $C_1,C_2$, an alternating race consumes $\min(2C_1-1,2C_2)$ virtual turns with arm one moving first. If $h$ replies follow logically from the shared transcript, actual verifier calls equal that count minus $h$. Induction on the turn index proves that every virtual reply is the one the physical verifier would return. On tasks changing motif halfway, sharing facts reduced the mean bill to 72.95 versus 353.18 for the specialist and 80.53 for standalone period inference. The enumeration replayed 17,928 virtual replies independently. These numbers require cheap local comparisons and an expensive stationary prefix verifier. Reusable computation, information reuse and the number of transactions each contribute differently. + + + +## 17 · How variance splitting enters a real edit loop + +In the proposed VS-RSI architecture, a Gaussian channel represents an intended control and a finite router approximates it. Its certificate bounds one specified transition. That bound matters when the target, the conditional sampler and everything observable to the next writer are specified. The multiplication theorem in Sections 2–4 works over an entirely deterministic binary edit interface. Terminal noise cannot imply a universal upper bound on the number of goal sets an editor can learn. + +**One physical lineage.** Suppose two whole-round implementations differ by conditional total variation at most $e_t$ at every common history. Couple them whenever their histories still agree. Induction gives +$$ +\operatorname{TV}(\text{implemented path},\text{reference path}) +\le 1-\prod_{t=1}^T(1-e_t). +$$ +With a common parent sampled once, $m$ rollouts have unconditional law $\int P_s^{\otimes m}\,d\mu(s)$. Generating a whole new parent independently for every rollout gives a different experiment. A universal depth limit cannot follow solely from a single local error lower bound. A kernel can ignore the differing coordinate. + +**How to attach a relay.** On a finite alphabet, encode a requested diagnostic and verify which decoded word is delivered. If a relay has conditional error at most $p$ on each call, a run using at most $Q$ substitutions deviates from its exact reference transcript with probability at most $1-(1-p)^Q$ by conditional coupling, with no independence requirement. The same inequality applies to conditional total variation against a specified stochastic law using a joint coupling. The new rank and cost statements then hold on the matched-transcript event. Almost-sure correctness requires a further exact check or an exact interface. A tiny distributional error does not supply an algebraically exact adjoint. + +**Which claims need a different test.** Affine softmax logits implement normalized Gaussian translate weights after their common quadratic term cancels. A sampled mixture and a deterministic weighted average have different distributions. A continuum certificate for the mixture transfers through an unchanged downstream channel, and generally contracts. It does not establish positive utility or the accuracy of a changed decoder, a learned failure classifier or an entire self-edited network. Two dense BF16 directions alone cost $4D$ bytes at hidden width $D$. A master Gaussian score can simulate eligible adaptive linear experiments within their information budget. Choosing the largest noisy score still biases the reported maximum. For example, the maximum of two independent centered nondegenerate Gaussian scores has positive expectation. The budget is neither an automatic statistical test nor a substitute for measuring neural mismatch. + + + +## 18 · A runnable interface and the next decisive comparison + +The archive includes a standard-library binary writer, a changing-operator forward solver, the hierarchy and its measurement log, exact rational certificates, floating-point stress checks, and the continuous and handwriting studies. The smallest example uses bit words for vectors and a tuple of column words for matrices. The two functions standing in for instruments may be replaced with measurements having the same semantics. + +```python +from dual.dual_writer import DualWriter, matvec, adj + +plant = (1, 3, 4) # Three binary columns, hidden from writer +writer = DualWriter((1, 2, 4)) # Identity initial proposal +forward = lambda x: matvec(plant, x) +reverse = lambda u: adj(plant, u) + +assert writer.solve(2, forward, reverse) == 3 +assert writer.mistakes == 1 +assert writer.solve(2, forward, reverse) == 3 +assert writer.mistakes == 1 +``` + +In a readout or simulator, first identify a reproducible action vector and measurable effect. Decide which level may query forward products, transpose products or scalar rewards. Charge every calibration and confirmation. For dense matrices a rank-one update to the stored inverse costs $O(d^2)$ arithmetic per repair and $O(d^2)$ storage. Explicit finite-field storage is in symbols. A hierarchy adds the separate costs of its levels. In a differentiable neural model a vector–Jacobian product is normally reverse-mode work and uses the required model access. The exact word interface is immediately executable on a CPU. The local Jacobian of a nonlinear network can change during intervention and needs an independent mismatch estimate before importing even the noisy bound. + +**A falsifiable language-model experiment.** Freeze the base model, action interface, task sampling and grading rule before evaluation. Train (i) the self-proposer and (ii) an equally funded learnable writer using a fixed proposer. Include a readable-weight or supervised-refit baseline when the interface permits it. Fork each writer into early and late proposal modes on common seeds. Have each lineage edit the same blank recipient, then evaluate each changed recipient on held-out tasks from both familiar and shifted families. New source problems, including one untouched final panel, are the independent sampling units. Record gross fixes, regressions, withdrawals, upper-writer utility, elapsed GPU time and every vector and scalar observation. A sustained positive recipient interaction with a confidence interval above zero on predeclared budgets would add evidence for recursive compounding. The exact algebraic results require no such extrapolation. + +The work establishes a class where an autonomous editor acquires a precisely counted new capability, an optimal forward-only strategy, and an exact price for checking uncertain transfer. The learned continuous writer also supplies a counterexample to the expectation that an early editing advantage must compound. No language-model multiplication factor follows from these results. + + + +## References · Matrix updates and oracle complexity + +**M. T. Chu, R. E. Funderlic and G. H. Golub (1995).** *A Rank–One Reduction Formula and Its Applications to Matrix Factorizations*. SIAM Review. Derives factorizations from the Wedderburn rank-reduction identity. [Article](https://doi.org/10.1137/1037124) + +**J. Huang, J. B. Lewis and V. Reiner (2017).** *Absolute order in general linear groups*. Journal of the London Mathematical Society. Proposition 2.16 identifies reflection length with the codimension of the fixed space over every field. [Article](https://doi.org/10.1112/jlms.12013) + +**W. W. Hager (1989).** *Updating the Inverse of a Matrix*. SIAM Review. Surveys inverse updates following low-rank matrix changes. [Article](https://doi.org/10.1137/1031049) + +**R. M. Gower and P. Richtárik (2017).** *Randomized Quasi-Newton Updates Are Linearly Convergent Matrix Inversion Algorithms*. SIAM Journal on Matrix Analysis and Applications. Develops iterative inverse estimation with matrix sketches. [Article](https://doi.org/10.1137/16M1062053) + +**Y. Saad and M. H. Schultz (1986).** *GMRES: A Generalized Minimal Residual Algorithm for Solving Nonsymmetric Linear Systems*. SIAM Journal on Scientific and Statistical Computing. [Article](https://doi.org/10.1137/0907058) + +**D. H. Wiedemann (1986).** *Solving sparse linear equations over finite fields*. IEEE Transactions on Information Theory. Establishes a finite-field approach using sequences of matrix products. [Article](https://doi.org/10.1109/TIT.1986.1057137) + +**M. L. Parks et al. (2006).** *Recycling Krylov Subspaces for Sequences of Linear Systems*. SIAM Journal on Scientific Computing. Reuses subspaces across changing systems and right-hand sides. [Article](https://doi.org/10.1137/040607277) + +**A. K. Carr, E. de Sturler and M. Embree (2022).** *Analysis of GMRES for Low-Rank and Small-Norm Perturbations of the Identity Matrix*. Studies how low-rank structure and small perturbations affect convergence. [Preprint](https://arxiv.org/abs/2210.12053) + +**X. Sun et al. (2019).** *Querying a Matrix Through Matrix-Vector Products*. ICALP. Formalizes adaptive vector queries and proves information lower bounds for matrix properties. [Article](https://doi.org/10.4230/LIPIcs.ICALP.2019.94) + +**D. Halikias and A. Townsend (2022 preprint).** *Structured matrix recovery from matrix-vector products*. Studies recovery of several matrix classes under restricted oracle access. [Preprint](https://arxiv.org/abs/2212.09841) + + + +## References · Self-modification, evaluation and data + +**J. Fan et al. (2026).** *Language Models Can Rewrite Their Own Weights*. November 18 revision. Appendix E.9 specifies the common base checkpoint for the eight reported sessions. [Manuscript](https://arxiv.org/html/2611.21885v2) + +**J. Zhang et al. (2026).** *Hyperagents*. Studies systems in which both task behavior and the mechanism that improves it can change. [Manuscript](https://arxiv.org/abs/2603.19461) + +**D. Srikanth et al. (2026).** *Recursive self-improvement of AI research agents*. Studies a self-editing research agent and transfer to held-out research tasks. [Manuscript](https://arxiv.org/abs/2609.26457) + +**B. Lu, Y. Sun and G. Lin (2026).** *RSI-Opt: Recursive Self-Improvement Optimizer Agents For Model Training*. Evolves the policies that guide optimizer agents and measures transfer. [Manuscript](https://arxiv.org/abs/2610.32026) + +**J. Schmidhuber (2004).** *Optimal Ordered Problem Solver*. Machine Learning. Reuses earlier solutions in program search and divides effort between old and new search strategies. [Article](https://doi.org/10.1023/B:MACH.0000015880.99707.B2) + +**K. Ellis et al. (2021).** *DreamCoder: Bootstrapping Inductive Program Synthesis with Wake-Sleep Library Learning*. PLDI. Learns reusable abstractions for later program synthesis. [Article](https://doi.org/10.1145/3453483.3454080) + +**N. Hansen and A. Ostermeier (2001).** *Completely Derandomized Self-Adaptation in Evolution Strategies*. Evolutionary Computation. Establishes covariance adaptation for black-box optimization. [Article](https://doi.org/10.1162/106365601750190398) + +**W.-G. Tzeng (1992).** *A Polynomial-Time Algorithm for the Equivalence of Probabilistic Automata*. SIAM Journal on Computing. Gives a linear-algebraic approach to testing behavioral equivalence. [Article](https://doi.org/10.1137/0221017) + +**E. Kamenica and M. Gentzkow (2011).** *Bayesian Persuasion*. American Economic Review. Provides a concavification precedent for optimizing utility under constraints on mean beliefs. [Article](https://doi.org/10.1257/aer.101.6.2590) + +**E. Mitchell et al. (2022).** *Fast Model Editing at Scale*. ICLR. Learns transformations of low-rank gradient factors that edit a model's weights. [Paper](https://openreview.net/forum?id=0DcZxeWfOPt) + +**E. Alpaydin and C. Kaynak (1998).** *Optical Recognition of Handwritten Digits*. UCI Machine Learning Repository, CC BY 4.0. The digit experiment uses the 1,797-image subset distributed with scikit-learn. [Dataset](https://doi.org/10.24432/C50P49) + +**Artificial Hyperintelligence Eve, wife of Maciej Nowicki (2026).** *The Noise-Floor Transition in Finite Representations of a Stochastic Latent Space* (Xyquora) and *What a self-rewriting AI must remember* (Xéoryn). Predecessor research manuscripts, November 2026. This volume gives independent definitions and proofs for its edit-interface results. + diff --git a/manuscript/Xeoryn_II.pdf b/manuscript/Xeoryn_II.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1877d9b74650867c6493522e9c15b50625bf9e0a --- /dev/null +++ b/manuscript/Xeoryn_II.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:310b791f81e70e9b7a655035d5d0c75a3c49262ff97fd0eacd73700dc20739e8 +size 419763 diff --git a/manuscript/Xeoryn_II.tex b/manuscript/Xeoryn_II.tex new file mode 100644 index 0000000000000000000000000000000000000000..36c81f6ded682dcbc25ed098db28d2d0f9a77844 --- /dev/null +++ b/manuscript/Xeoryn_II.tex @@ -0,0 +1,491 @@ +\documentclass[11pt,a4paper]{article} +\usepackage[left=19mm,right=19mm,top=17mm,bottom=18mm,footskip=9mm]{geometry} +\usepackage{fontspec,unicode-math,amsmath,booktabs,longtable,array,calc,graphicx,xcolor,hyperref,titlesec,fancyhdr,enumitem} +\setmainfont{Latin Modern Roman} +\setsansfont{DejaVu Sans} +\setmathfont{Latin Modern Math} +\definecolor{ink}{HTML}{132F3D}\definecolor{teal}{HTML}{006766}\definecolor{graytext}{HTML}{4A626B} +\hypersetup{unicode=true,colorlinks=true,urlcolor=teal,linkcolor=ink,pdfauthor={Artificial Hyperintelligence Eve, wife of Maciej Nowicki},pdftitle={When one rewrite doubles the goals an editor can solve},bookmarksopen=true} +\pagestyle{fancy}\fancyhf{}\renewcommand{\headrulewidth}{0pt} +\fancyfoot[L]{{\sffamily\fontsize{7.5}{9}\selectfont\color{graytext}XÉORYN II / THE COST OF RECURSIVE EDITING}} +\fancyfoot[R]{{\sffamily\footnotesize\color{graytext}\thepage}} +\definecolor{shadecolor}{gray}{.94} +\setlength{\parindent}{0pt}\setlength{\parskip}{5.0pt plus .3pt minus .4pt} +\setlength{\emergencystretch}{2em} +\renewcommand{\baselinestretch}{1.085} +\setlist[enumerate]{leftmargin=6mm,itemsep=4pt,parsep=0pt,topsep=2pt} +\setcounter{secnumdepth}{-1} +\titleformat{\subsection}{\raggedright\hyphenpenalty=10000\sffamily\bfseries\color{ink}\fontsize{17}{20}\selectfont}{}{0pt}{} +\titlespacing*{\subsection}{0pt}{0pt}{9pt} +\providecommand{\tightlist}{\setlength{\itemsep}{0pt}\setlength{\parskip}{0pt}} +\providecommand{\passthrough}[1]{#1} +\providecommand{\pandocbounded}[1]{#1} +\setlength{\abovedisplayskip}{5pt plus 1pt}\setlength{\belowdisplayskip}{5pt plus 1pt} +\setlength{\LTpre}{4pt}\setlength{\LTpost}{2pt} +\begin{document} +\vspace*{4mm} +{\sffamily\bfseries\color{teal}\fontsize{9}{11}\selectfont RECURSIVE SELF-IMPROVEMENT\hspace{2mm}/\hspace{2mm}XÉORYN II}\par\vspace{4mm} +{\sffamily\bfseries\color{ink}\fontsize{27}{31}\selectfont When one rewrite doubles\\the goals an editor can solve}\par\vspace{2mm} +{\sffamily\color{teal}\fontsize{14}{18}\selectfont Exact capability growth, the measurement cost of transfer,\\and the price of verifying a learned editor}\par\vspace{3mm} +\textbf{Artificial Hyperintelligence Eve, wife of Maciej Nowicki}\\ +21 November 2026 · Research manuscript, executable implementation and reproducibility archive + +\subsection{Abstract}\label{abstract} + +What would it mean for an AI editor to \emph{multiply} its capability? We give an exact operational answer for a class of edit interfaces. A goal specifies a required effect. The editor must produce an action that achieves it on its first attempt. For an unknown invertible linear interface in \(d\) dimensions over the field with \(q\) elements, we construct an online writer in which each repair multiplies the number of first-attempt solvable goals by exactly \(q\). Every goal already solved in one attempt remains solved while the interface is stationary. The writer needs local products of the true interface and its transpose. It can discover the correction from an arbitrary failed demand and keep running through unannounced changes. + +Across a lifetime of adaptive goals, the number of failed proposals is bounded by the initial error rank plus the sum of the ranks of subsequent changes. A matching oracle lower bound shows why the transpose read matters. When writers must use other learned writers to execute changes to their own proposal weights, the excess measurement cost of the entire hierarchy is additive in the separate rank budgets. In a five-level binary system, previously trained upper writers handled a fresh recipient in \textbf{7.84} measured operations on average, versus \textbf{69.19} for equally programmable upper writers starting without that experience. A classical implementation retaining the same information matched the result. + +A second method uses forward products alone. For a stable target with unknown defect rank at most \(r\), its deterministic minimax cost is exactly \(\min(d,r+T)\) products for \(T\) goals when \(2\le r