Searcharxiv⌕ Search

arXiv subjects

Tong Niu

Publications and source records attributed to Tong Niu.

At least 19 recordsLinked to original sources

EgoLink: Egocentric Language-Vision Interactive Network Knowledge Challenge Summary

The EgoLink Challenge at ACM Multimedia 2026 benchmarks egocentric multimodal intelligence linking first-person perception, social reasoning, and interactive action. Unlike benchmarks centered on visible objects and activities, EgoLink evaluates whether models can interpret emotional and social cues, resolve visually grounded references, and convert user requests into verified tool-mediated actions. Its two tracks cover social reasoning in egocentric video and interactive agents that complete service tasks through dialogue, tool use, and closed-loop execution. Together, they test whether perception, reasoning, and decision-making remain coupled under both offline question answering and state-changing interaction. We highlight a multi-stage Track 1 data pipeline for multiple-choice question (MCQ) generation, chain-of-thought generation, automatic quality grading, and video-contribution filtering. We summarize the task design, evaluation metrics, rules, leaderboards, and representative solutions. The benchmark combines objective answer scoring with process- and result-based evaluation, enabling analysis of both final correctness and execution quality. The best Track 1 system achieved 0.9263 accuracy, while the best Track 2 system achieved 0.844 joint success and 0.916 micro accuracy. Participant reports reveal a shared principle: strong systems make intermediate evidence explicit through candidate verification, option-robust reasoning, scenario-aware planning, tool canonicalization, and state-aware correction before answering or acting. These findings highlight the importance of process verification and evidence-grounded reasoning for reliable egocentric multimodal intelligence.

cs.CV↗

ToolSearcher: Optimizing Tool Selection at Scale via Reinforcement Learning

Large language models (LLMs) excel at natural language processing but struggle to interact with external environments. Tool learning provides a promising way to extend LLMs into actionable agents, where tool selection is a critical prerequisite for successful tool use. Existing work often assumes a small or predefined set of tools, leaving large-scale tool selection underexplored. Real-world repositories contain a vast and diverse array of tools, making it difficult for LLMs to effectively search, distinguish, and compose tools under context-length constraints. We identify large-scale tool selection as a new challenge for agentic reinforcement learning, highlighting that existing RL methods for knowledge-based question answering are inadequate for selecting tools while considering compatibility. To address this challenge, we propose ToolSearcher, a novel RL framework for effective multi-turn search and fine-grained optimization in large-scale tool selection. Specifically, we introduce category-constrained tool discrimination to improve the model's ability to distinguish functionally similar tools, event-level search modeling to explicitly optimize the discovery of target tools during multi-turn search, and trajectory-aligned credit allocation to provide fine-grained reward signals for different stages of the search-selection process. Extensive experiments on large-scale tool selection benchmarks demonstrate that ToolSearcher consistently outperforms a set of strong baselines in challenging settings involving iterative search and complex tool composition.

cs.CL↗

Non-adjacent triples of cells in a triangular grid: proofs of three conjectured generating functions

Let $\Tn$ be the triangular grid of side $n$, with its $n^2$ unit cells, and let two cells be side-adjacent if they share an edge and point-adjacent if they share a vertex. We count the orbits of $3$-element sets of pairwise non-adjacent cells under the rotation group $C_3$ and under the full symmetry group $D_3$ of $\Tn$. Three of the four resulting sequences are OEIS A397730, A397731 and A397732, for each of which the only formula on record is a generating function conjectured by Sean A.\ Irvine on 14 July 2026. We prove all three conjectures. The route is Burnside's lemma, together with exact lattice counts for each fixed-point term. The side-adjacency graph turns out to be bipartite, so its triangle count vanishes, and the triangles of the point-adjacency graph have a clean structural description, which gives that count too. Along the way we obtain the first explicit closed forms for the three sequences, each a quasi-polynomial of degree $6$ and period $3$, $6$, $3$ respectively. The same machinery settles the fourth member of the family, whose closed form has been recorded at A289230 without proof since 2017, and the labeled count A289224 that feeds it.

math.CO↗

EgoBench: An Interactive Egocentric Multimodal Benchmark for Tool-Using Agents

As AI agents increasingly operate in open, real-world environments, they require a deep synergy of multimodal perception, tool invocation with multi-hop reasoning, and dynamic interaction with users. However, existing benchmarks fail to jointly evaluate these capabilities due to challenges in designing strictly coupled multi-capability tasks, simulating natural and task-constrained user feedback, and ensuring objective evaluation of dynamic interaction. To bridge this gap, we introduce EgoBench, the first interactive multimodal benchmark for tool-using agents. EgoBench comprises 1,590 egocentric-video-grounded tasks covering five daily scenarios, along with a user-agent-tool interactive environment for evaluation. We implement a three-stage synergistic pipeline through which each task is designed to enforce the joint application of visual perception and tool-augmented multi-hop reasoning. We additionally develop a multi-agent simulated user to evaluate agents' interaction capabilities, which generates high-fidelity, task-aligned responses to agents. Furthermore, we establish a deterministic joint validation framework that guarantees objective assessment through process-based and result-based equivalence. Benchmarking eight video-MLLM agents on EgoBench reveals a severe performance ceiling: the best-performing model achieves only a 34.95% average Joint Success Rate across the three interaction modes. Finally, we conduct a multi-dimensional error analysis to disentangle failure modes, exposing capability bottlenecks for advancing future AI agents.

cs.AI↗

The first descent in a standard Young tableau of shape $(n,n,n)$

Let $a(n)$ be the number of standard Young tableaux of shape $(n,n,n)$ whose entry in row $2$, column $1$ is odd; equivalently, the number of those whose first descent is even. This is entry A011553 of the On-Line Encyclopedia of Integer Sequences, contributed in 1996, and after thirty years of curation it carries no formula. We supply one, $a(n) = 8\,\bigl(n!\,(n+2)!\bigr)^{-1}\sum_{m=1}^{\lfloor n/2\rfloor} m(m+1)(3n-2m-1)!/(n-2m)!$, and use it to settle both of the conjectures the entry records. A creative telescoping certificate shows that $a$ satisfies a linear recurrence of order two with polynomial coefficients; the order-three recurrence conjectured by R. J. Mathar in 2023 is a left multiple of it, with explicit cofactor $(4S^{-1}-3)/(7n-9)$. We also prove $a(n)\sim 3^{3n+7/2}/(64πn^{4})$, the asymptotic conjectured by V. Kotesovec in 2014. The second proof gives slightly more than the conjecture asks: the position of the first descent has a limiting distribution, the probability that the $(2,1)$ entry equals $r+1$ tending to $r(r+2)/3^{\,r+1}$. Summing the even terms, a uniformly random tableau of shape $(n,n,n)$ has an odd $(2,1)$ entry with probability tending to $27/64$, whereas for a uniformly random tableau of $n$ cells of unrestricted shape the corresponding limit is $1/e$.

math.CO↗

A short proof of Mathar's 2021 recurrence conjecture for the Lehmer-Comtet diagonal A045406

For OEIS sequence A045406, the column-2 diagonal of the Lehmer-Comtet triangle A008296, R. J. Mathar contributed in September 2021 the conjectured order-2 P-recursive recurrence \[ a(n) + (2n-7)\,a(n-1) + (n-4)^{2}\,a(n-2) \;=\; 0,\qquad n \ge 5. \] We give a short proof. Detlefs's harmonic-number closed form $a(n) = (-1)^n (2 H_{n-3} - 3)(n-3)!$ for $n \ge 3$ collapses the left-hand side, after factoring out $(-1)^n (n-5)! (n-4)$, to a polynomial identity in $n$ with coefficient $H_{n-4}$. The $H_{n-4}$-coefficient simplifies to $(n-3) - (2n-7) + (n-4) = 0$ (using $H_{n-3} = H_{n-4} + 1/(n-3)$ and $H_{n-5} = H_{n-4} - 1/(n-4)$); the constant remainder is $0$ for the same reason. The supplementary archive contains a SymPy script verifying both pieces symbolically, the e.g.f.\ expansion against the harmonic closed form, and Mathar's recurrence numerically for $n = 5, \ldots, 5000$.

math.CO↗

An exact small-$n$ computation of the minimum 2-coloring discrepancy of $K_n^{(3)}$

For an integer $r \ge 2$ and an order $n \equiv 1, 3 \pmod{6}$, write $δ_r(n)$ for the minimum, over all $r$-colourings $χ: \binom{[n]}{3} \to [r]$, of $\max_{\mathcal{S}} \mathrm{disc}(\mathcal{S}, χ)$, where the maximum is over labelled Steiner triple systems $\mathcal{S}$ of order $n$ and $\mathrm{disc}(\mathcal{S}, χ) = \max_c |\#\{T \in \mathcal{S} : χ(T) = c\} - |\mathcal{S}|/r|$. Following Gishboliner, Glock, and Sgueglia \cite{GishbolinerGlockSgueglia2025}, the bulk of the recent work on this quantity has been on lower bounds for $r \ge 3$ (proving $δ_r(n) = Ω(n^2)$) and on structural characterisation of the low-discrepancy 2-colourings. We give three small computational contributions in the small-$n$ regime $n \in \{7, 9, 13, 15, 19, 21\}$: An exact value of $δ_2(n)$ for each such $n$, matching the formula $δ_2(n) = \min_{x \in [0, n] \cap \mathbb{Z}} |x(n-x)/2 - n(n-1)/12|$ obtained by optimising the GGS Example 1.1 family. Rigorous for $n \in \{7, 9\}$ via exhaustive search over labelled STSs ($30$ resp. $840$ systems) and over all $2$-colourings; computational for $n \in \{13, 15, 19, 21\}$ by simulated-annealing search; A wide near-optimal basin: at $n = 9$, every two-colour-flip neighbour of the optimal Example~1.1 colouring that maintains discrepancy $1.0$ exists; about $34\%$ of two-flip perturbations preserve optimality; Random-colouring statistics for $r \in \{2, 3, 4\}$: $\langle\max_{\mathcal{S}}\mathrm{disc}\rangle$ grows linearly in $n$, in agreement with a heuristic Gaussian estimate $n / \sqrt{6r} \cdot \sqrt{2 \log K}$ over $K$ sampled labellings; the typical-case discrepancy is far below the GGS worst-case $Ω(n^2)$. We additionally state a conjectural exact formula for $δ_2(n)$ that holds for every $n \equiv 1, 3 \pmod{6}$.

math.CO↗

The pebbling number of Fibonacci cubes

The $n$-th Fibonacci cube $Γ_n$ is the subgraph of the hypercube $Q_n$ induced by binary strings with no two consecutive ones. We determine $π(Γ_n) = 2^n$ for $n \le 6$, so the pebbling number of $Γ_n$ equals that of the ambient hypercube $Q_n$ despite $Γ_n$ having far fewer vertices. The lower bound is a standard potential argument. For the upper bound, the Weight Function Lemma yields $2^n+1$ -- one too many -- so we close the gap by exhaustive MILP verification. We conjecture $π(Γ_n) = 2^n$ for all $n$.

math.CO↗

Parity vectors and paradoxical sequences in the accelerated Collatz map

This note studies parity vectors and paradoxical sequences in the accelerated Collatz iteration $T(n) = (3n+1)/2$ for $n$ odd, $T(n) = n/2$ for $n$ even. Building on Rozier and Terracol (arXiv:2502.00948, 2025), Terras (1976), Lagarias (1985), and Tao (2019), we prove three theorems and add one numerical observation. The first is a sharp finitary form of Terras's parity-vector density; the second is a closed-form analytic count of paradoxical $Ω_k(n)$ for each fixed length $k$. The third is a density-zero theorem for bounded-length paradoxical sequences with explicit constant. As for the numerical piece, among the seven $(j, q)$ pairs that show up in the Rozier-Terracol enumeration with first term $n \le 10^9$, every paradoxical reduced ratio $q/j$ turns out to be a left convergent, a left semiconvergent, or a Stern-Brocot mediant of adjacent convergents/semiconvergents of $\log_3 2$. The three theorems are unconditional. The fourth observation is verified for $n \le 10^7$ and conjectured for all $n$. We make no claim toward the Collatz conjecture or Terras's coefficient-stopping-time conjecture.

math.NT↗

Toward a Formally Verified Optimality Certificate for OGR(29): A SAT-Encoding Methods Note with Small-Case Demos

A Golomb ruler of order~$n$ is an integer set $\{a_0<a_1<\cdots<a_{n-1}\}$ whose $\binom{n}{2}$ pairwise differences $a_j-a_i$ ($i<j$) are all distinct. The optimal Golomb ruler problem asks for $\mathrm{OGR}(n)=\min\{a_{n-1}-a_0\}$ and is a classical combinatorial benchmark. The values $\mathrm{OGR}(2),\dots,\mathrm{OGR}(28)$ are settled through distributed volunteer search (the Distributed.net OGR project); $\mathrm{OGR}(29)$ is in active computation, with verification expected in late 2026 or early 2027. The recent upper bound $\mathrm{OGR}(29)\le 757$ of Lee, Park, and Kim (arXiv:2510.0122, October~2025) tightens the search window. This note describes a compact CNF encoding of the decision problem $\mathrm{GR}(n,L)$ (``is there a Golomb ruler of order~$n$ with length exactly~$L$?'') with $O(n^2 L)$ clauses, together with a small-case verification sweep that emits machine-checkable LRAT certificates of optimality for $\mathrm{OGR}(n)$ at $n\le 12$. We then give closed-form encoding-size estimates for $\mathrm{OGR}(29, L=757)$ and propose a cube-and-conquer decomposition aimed at a Mallob-style parallel run on commodity multi-core hardware. The certificate pipeline (CaDiCaL with --lrat=true, a structural sanity-check, then formal validation by drat-trim or cake\_lpr) is end-to-end. A formally verified optimality proof for any single $\mathrm{OGR}(n)$ value beyond the trivial $n\le 5$ would be a first in the field.

cs.DC↗

Three short proofs of Mathar's 2014 conjecture for OEIS A002627

For the OEIS sequence A002627, defined by the inhomogeneous first-order recurrence $a(n) = n\,a(n-1) + 1$ with $a(0) = 0$, R.~J.~Mathar recorded in February 2014 the conjectured second-order homogeneous recurrence \[ a(n) - (n+1)\,a(n-1) + (n-1)\,a(n-2) = 0, \qquad n \ge 2, \] which has remained marked as a conjecture on the OEIS for over a decade. We give three short proofs. The first is two lines: subtract the defining recurrence at adjacent indices and the constant cancels (we call this homogenisation). The second reads off the same relation from the exponential generating function $F(x) = (e^x-1)/(1-x)$. The third is a Pascal-rule telescoping on the binomial-sum form $a(m) = \sum_{k=0}^{m-1} k!\binom{m}{k}$. All three derivations are elementary, requiring nothing beyond undergraduate techniques. We remark that the same homogenisation trick clears an entire class of ``Conjecture: \dots'' entries on the OEIS, namely sequences satisfying $a(n) = p(n)\,a(n-1) + q(n)$ with simple $q$.

math.CO↗

An explicit algebraic generating function for OEIS A348410

For the OEIS sequence A348410, P. Bala recorded in February 2022 two equivalent closed forms, $a(n) = [x^{n}] ((1-x)(1-x^2))^{-n}$ and a single-index binomial sum. R. J. Mathar (October 2021) and V. Kotesovec (November 2021) each contributed a conjectured P-recursive recurrence -- Mathar's of order $4$, Kotesovec's of order $2$. We apply Lagrange-Bürmann inversion to Bala's $[x^n]$ form to derive the parametric expression $A(t) = (1 - y^2)/(1 - y - 4 y^2)$, where $y = y(t)$ is implicit by $y(1-y)^2(1+y) = t$. Eliminating $y$ via resultant gives the explicit algebraic equation $P(t, A) = 0$ of degree $4$ in $A$ and degree $2$ in $t$. As an immediate corollary (Stanley's classical algebraic-implies-D-finite theorem), $A(t)$ is D-finite. Mathar's and Kotesovec's specific recurrences are not directly proven here; we only verify Kotesovec's order-$2$ recurrence numerically for $n = 3, \ldots, 1000$ and observe that an explicit ODE-and-recurrence extraction from $P(t, A) = 0$ via the standard Bostan-Chyzak-Salvy algebraic-to-holonomic procedure would close both conjectures. The supplementary archive contains a SymPy script which derives $P(t, A)$ and checks the numerical evidence.

math.CO↗

Size-4 Counterexamples to the Sidon-Extension Conjecture

A finite set $S \subset \mathbb{Z}$ is a Sidon set if its pairwise differences are distinct. Recall that a perfect difference set (PDS) of order $n$ is a set $B \subset \mathbb{Z}_v$ ($v = n^2 - n + 1$) of size $n$ such that every nonzero residue arises exactly once as a difference of two elements of $B$. Erdős's \$1000 conjecture -- that every finite Sidon set extends to a finite PDS -- was disproved by Alexeev and Mixon (arXiv:2510.19804, October 2025), via the size-5 counterexamples $\{1,2,4,8,13\}$ and Hall's earlier $\{1,3,9,10,13\}$; they then asked: what is the smallest size $s$ of a non-extending Sidon set? The trivial bounds give $3 \le s \le 5$. Our evidence points to $s = 4$. We exhibit two integer Sidon sets, \[ A = \{0, 1, 3, 11\}, \qquad B = \{0, 1, 4, 11\}, \] together with the apparent infinite family of dilations $kA$, $kB$ and their reflections, all of which fail to extend for every prime power $q \le 317$ via the Singer affine-orbit check (rigorous under Hall's 1947 uniqueness for Desarguesian cyclic planes through $q \le 40$ and under the prime-power conjecture beyond that), and unconditionally for every modulus $v \le 133$ via brute-force depth-first search. We also report the exact density $N_{\text{ne}}(N) = 4 \lfloor N / 11 \rfloor$ of non-extending size-4 Sidon sets in $[0, N]$ for $N \le 50$ -- the match is exact, which suggests the $kA, kB$ family is complete in this range. A complete proof, perhaps in the spirit of Alexeev--Mixon's polarity argument or via a multiplier descent, remains open.

math.CO↗

Graceful labelings of spiders with three-edge legs and pendant leaves at the center

A graph $G$ on $m$ edges is graceful if there is an injection $f : V(G) \to \{0, 1, \ldots, m\}$ whose induced edge labels $\{|f(u) - f(v)| : uv \in E(G)\}$ are exactly $\{1, 2, \ldots, m\}$. Ringel and Kotzig conjectured in 1964 that every tree is graceful. A computer check has confirmed this for all trees on at most 35 vertices (Fang 2010), but no general proof is known. Here we exhibit an infinite family of trees that escapes the named spider results of Bahls--Lake--Wertheim, Panpa--Poomsa-ard, and Panpa--Imnang--Wasuanankul: the family $T(k, m)$ of spiders with $k$ legs of length $3$ together with $m$ pendant leaves at the centre. We prove every such tree is graceful for all $k \ge 1$ and $m \ge 0$. The argument splits into two short lemmas. The first is a pendant-extension lemma that applies whenever the underlying graceful labeling sends the centre to $0$; the second is the base case, namely that $S_{k, 3}$ admits exactly such an apex-zero labeling, a fact already implicit in Bahls--Lake--Wertheim (2010). What is new is the explicit identification of $T(k, m)$ and the observation that the family is closed under both pendant addition and the apex-zero condition, so it includes infinitely many trees not handled by the named theorems above.

math.CO↗

Improved Upper Bounds on the Pebbling Numbers of the Blanuša Snarks

The two Blanuša snarks $B_1$ and $B_2$ are 3-regular graphs on 18 vertices. Dantas, Lordelo, Niedermaier and Nogueira (Discrete Appl. Math. 361, 2025, pp. 336-346) established the first systematic bounds $23 \le π(B_i) \le 34$ for $i=1,2$. Bridi, Marquezino and Figueiredo (arXiv:2505.16050, 2025) then sharpened the upper side to $π(B_1) \le 31$ and $π(B_2) \le 30$ via a Weight Function Lemma heuristic. We push the upper bounds further to $π(B_1) \le 28$ and $π(B_2) \le 29$. The route is again Hurlbert's Weight Function Lemma, but applied one automorphism orbit at a time, with optimal weight functions coming from a linear program over a corpus of roughly $30{,}000$ rooted-subtree strategies per target. For the lower bound $π(B_i) \ge 23$ we re-derive the witnesses of Dantas et al. and re-verify them with two independent oracles: an exhaustive forward state-space search, and a sound-and-complete MILP encoding whose acyclicity constraint is motivated by the Milans-Clark No-Cycle Lemma. The interval for $B_1$ shrinks from $[23, 31]$ to $[23, 28]$, and for $B_2$ from $[23, 30]$ to $[23, 29]$.

math.CO↗

A short proof of Mathar's 2013 recurrence conjecture for the reversible-binary-string sequence A032123

For the OEIS sequence A032123, the number of length-$2n$ black-and-white strings with $n$ black beads, considered up to reversal, R. J. Mathar contributed in November 2013 the conjectured order-5 P-recursive recurrence \[ \begin{aligned} &n(n-1)\,a(n) - 2(n-1)(3n-4)\,a(n-1) + 4(2n^{2}-14n+19)\,a(n-2) &\qquad + 8(n^{2}+5n-19)\,a(n-3) - 16(n-3)(3n-10)\,a(n-4) &\qquad + 32(n-4)(2n-9)\,a(n-5) \;=\; 0, \qquad n \ge 6. \end{aligned} \] We give a short proof. Burnside's lemma applied to the reversal action gives the closed form $a(n) = \tfrac{1}{2}\bigl(\binom{2n}{n} + [n \text{ even}]\binom{n}{n/2}\bigr)$; the two summands satisfy elementary recurrences of order $1$ and $2$ respectively; and Mathar's order-5 operator, applied to each summand separately, reduces to a polynomial identity that simplifies to zero after a brief calculation. The supplementary archive includes a SymPy script which verifies the polynomial identities symbolically and checks Mathar's recurrence numerically for $n = 6, \ldots, 5000$.

math.CO↗

A short proof of Mathar's 2020 recurrence conjecture for the generalized-Stirling sequence A001711

For the OEIS sequence A001711, contributed by N. J. A. Sloane long before the on-line era and identified there as the diagonal $T(n+4, 4)$ of a generalized-Stirling triangle, R. J. Mathar contributed in February 2020 the conjectured order-2 P-recursive recurrence \[ a(n) - (2n+5)\,a(n-1) + (n+2)^{2}\,a(n-2) \;=\; 0,\qquad n \ge 2. \] We give a one-page proof. Detlefs's harmonic-number closed form $a(n) = \tfrac{1}{4}(n+3)!\,(2 H_{n+3} - 3)$ collapses the left-hand side, after dividing through by $(n+1)!/4$, to a polynomial identity of $n$ with coefficient $H_{n+2}$. The harmonic-number coefficient simplifies to $(n+3) - (2n+5) + (n+2) = 0$ (using $H_{n+3} = H_{n+2} + \tfrac{1}{n+3}$ and $H_{n+1} = H_{n+2} - \tfrac{1}{n+2}$); the constant remainder is $-3 \cdot 0 = 0$ for the same reason. The supplementary archive contains a SymPy script verifying both pieces symbolically, the e.g.f.\ expansion against the harmonic closed form, and Mathar's recurrence numerically for $n = 2, \ldots, 5000$.

math.CO↗

The vertex Folkman number $F_v(3,3;5)$ equals~$8$

The vertex Folkman number $F_v(s,t;k)$ is the smallest $n$ for which there exists a $K_k$-free graph on $n$ vertices whose vertices cannot be $2$-colored without producing a monochromatic copy of $K_s$ or $K_t$. We show $F_v(3,3;5)=8$. The witness is the cone $K_1 \vee \overline{C_7}$, a single universal vertex joined to the complement of a $7$-cycle. That this graph is $K_5$-free and arrows $(3,3)^v$ follows from a short independence-number argument. The matching lower bound -- no $K_5$-free graph on $7$ or fewer vertices works -- comes from exhaustive enumeration via nauty and a SAT check using Glucose\,4. The appendix has a self-contained Python script for verification.

math.CO↗